Blockchain-based Federated Learning: Models, Methods and Applications

Source: Acta Automatica Sinica

Authors: Li Cheng, Yuan Yong, Zheng Zhiyong, Yang Dong, Wang Feiyue

Note: If you can't display the full article, you can V search for "Artificial Intelligence Technology and Consulting" to view the full article

Artificial Intelligence, Big Data, Multimodal Large Models, Computer Vision, Natural Language Processing, Digital Twins, Deep Reinforcement Learning······ You can also add V "Artificial Intelligence Technology and Consulting" to sign up for the course

summary

In recent years, human society has rapidly entered the era of big data, and data security and privacy protection have become the key issues in the development of big data ecology and related digital economy. Federated learning, as a new paradigm of distributed machine learning, is committed to training global models from distributed local datasets while protecting data privacy, so it has been widely and deeply studied. However, the technical challenges faced by the federated learning system, such as centralized architecture, incentive mechanism design and system security, still need to be further studied, and blockchain is considered to be an effective solution to deal with these challenges, and has been successfully applied to many research and practice scenarios of federated learning. On the basis of systematically combing the current research results of blockchain and federated learning, this paper proposes a blockchain-enabled federated learning (BeFL) conceptual model, expounds some key technologies, research problems and current research progress, discusses the application scenarios and key issues to be further studied in this field, and discusses the potential directions of future development. It is committed to providing useful reference and reference for building a decentralized, secure and credible data ecological infrastructure and promoting the development of digital economy and related industries.

keyword

Blockchain / Federated Learning / Smart Contract / Machine Learning / Privacy Protection

With the rapid development of information and intelligent technology, the data generated by the Internet and Internet of Things devices has shown an explosive growth trend. According to the International Data Corporation (IDC), the global data scale forecast shows that by 2025, the global data will reach 175 zettabytes (zettabytes), of which 90 zettabytes will come from IoT devices, and the number of data interaction users will increase from 5 billion in 2018 to 6 billion. Data has become a new, more important and effective factor of production. For example, the new generation of artificial intelligence technologies represented by deep learning usually requires a large amount of data to train an ideal model, in order to improve the performance and efficiency of intelligent systems.

However, while data plays an important role, its collection and use will also be related to personal safety and national security, so domestic and foreign laws and regulations on data privacy protection are becoming increasingly stringent: in 2018, the European Union issued the General Data Protection Regulation (GDPR); In 2021, China successively promulgated the Data Security Law and the Personal Information Protection Law. Although these laws and regulations help to protect data security and privacy, they also limit data circulation and value creation to a certain extent, forcing data to be scattered in disconnected data silos due to factors such as security, privacy or geographical location. Therefore, under the premise of ensuring privacy and security, how to promote data circulation and sharing, and enhance the efficiency of collaboration and cooperation between institutions is a common concern in academia and industry.

Federated learning is a new paradigm of distributed machine learning that has emerged in recent years, which can realize that the private data of each institution can be built without disclosing the underlying raw data or its encrypted form data without disclosing the underlying raw data or its encrypted form data. Because the data itself is not moved, it effectively reduces the risk of privacy leakage and data compliance, so federated learning has received extensive and in-depth research attention in recent years[1−3]. In terms of theoretical construction, the research on federated learning can be traced back to 1996, when Cheung et al. [4] realized association rule mining in distributed databases for the first time, laying a theoretical and methodological foundation for federated learning. In 2016, Google officially proposed federated learning technology and used it to achieve input method optimization[5]. In 2017, Tan et al. [6] proposed a theoretical system of distant domain transfer learning, combined transfer learning with federated learning, and proposed a technical framework for federated transfer learning in 2020 to solve the problem of data silos[7].

In recent years, three forms of federated learning have been developed and derived in industrial practice, namely, horizontal federated learning (sample dimension), vertical federated learning (feature dimension), and federated transfer learning[8−9]. In terms of application architecture, horizontal federated learning is divided into client-central coordinator architecture and peer-to-peer network architecture, and vertical federated learning usually involves a third party with the participation of two parties and assumes that there is a third party with mutual trust as a coordinator to collaborate with both parties for privacy computation. Federated transfer learning is based on a privacy-preserving distributed machine learning architecture, combined with traditional transfer learning methods to achieve knowledge transfer. In terms of application scenarios, according to the data distribution characteristics of the participants, horizontal federated learning is suitable for application scenarios where there is more overlap of user features and less overlap of user samples. Vertical federation is suitable for application scenarios with little overlap of user features and many overlapping user samples. Federated transfer learning is suitable for application scenarios where user features and user samples overlap less.

At present, federated learning still faces many challenges[10−11]. 1) At the infrastructure level, the underlying network topology of mainstream federated learning relies on a centralized server to process the parameters uploaded by each node, and once a single point of failure occurs on the server, the entire system will be paralyzed. At the same time, with the increase of nodes participating in the training, the network load of the centralized server will also increase accordingly, which will reduce the efficiency of system joint training. On the other hand, although the peer-to-peer (P2P) network architecture of the peer-to-peer learning network enables the participating nodes to pass encrypted gradients or model parameters to each other without going through a central server for aggregation, thus effectively alleviating the challenge of single point of failure, the whole system lacks a unified scheduling mechanism to coordinate multiple participants for federated computing. It can be seen that the traditional federated learning network topology restricts the robustness and efficiency of the whole system. 2) At the data security level, although federated learning can protect the data from being stolen or tampered with during transmission by using encryption algorithms, it cannot prevent adversaries from inferring data information by analyzing model parameters or outputs. For example, an adversary can take a member inference attack and a feature inference attack, by observing the changes in model parameters, to infer whether the participant's training data contains some specific samples or attributes. Alternatively, the adversary can adopt a refactoring attack and a model inversion attack, by constructing a special input, observing the output of the model, and deducing the content of the training data in reverse. At the same time, federated learning mainly focuses on the data computation process, and there are few effective methods for input data screening and output data or parameter checking[12−13], so it is not perfect in terms of data security and integrity. 3) At the incentive mechanism level, users participating in joint training need to contribute their computing resources and private data to train a global model shared by all parties; In practical applications, there are often great differences in the computing resources and data quality of each user, so the participants with superior resources and high-quality data usually lack the motivation to participate in federated learning in order to maintain their industry advantages. Therefore, federated learning needs an effective incentive mechanism to motivate users to participate. 4) At the node trust level, the premise of federated learning is that the nodes participating in the training jointly trust the central server to do parameter processing. In a dynamic and open network environment, it is difficult to find a central node that all participants can trust. Therefore, how to build trust among participants is the key to improve the effectiveness of federated learning.

The technical advantages of blockchain technology, such as decentralized infrastructure, user identity security authentication mechanism, automatic incentive distribution mechanism and immutability of block data, are expected to provide solutions to the challenges of federated learning. The basic workflow of the blockchain system is that distributed blockchain nodes participate in a pre-set consensus process through the P2P network to complete the verification of transaction or transaction data, and encapsulate these data in a chain block structure, so as to maintain the same data ledger among consensus nodes. The consensus process is usually for each node to obtain the accounting power according to a predefined consensus mechanism (e.g., competitive bookkeeping based on hashrate or equity, rotational bookkeeping based on a specific order, etc.), and the winning node packages all the data generated in the current time period, encapsulates it into a new block, and links it to the main chain in chronological order. At the same time, the blockchain system may issue a certain number of tokens to reward the winning node and incentivize other nodes to continue to participate in the data consensus process[14].

As a new paradigm of distributed computing, blockchain can improve federated learning from the following four aspects. 1) The blockchain network adopts a fully decentralized (or weakly centralized) P2P network topology, which provides a suitable architecture for the aggregation of distributed models of federated learning, and improves the computational elasticity, system integrity and fault tolerance. 2) The identity authentication and permission management mechanisms of the blockchain system can improve the security of the federated learning system; 3) Blockchain can automatically manage multi-round federated learning tasks for different sets of devices through custom smart contracts, and at the same time, more users can be incentivized to participate in the co-construction of the ecosystem through cryptocurrency; 4) With the support of the underlying distributed consensus protocol, the blockchain can ensure the fairness and impartiality of the federated learning process, and help establish trust between user nodes participating in the training.

Therefore, in recent years, the combination of blockchain technology and federated learning has become a new trend to ensure data security and privacy and build a new infrastructure for the circulation of data elements. The two are mutually beneficial, and will form a more complete solution in motivating federated learning participants to conduct collaborative data training while ensuring data privacy and security[15−17].

The potential advantages of blockchain-enabled federated learning (BeFL) architecture include the following three aspects. 1) Similar cooperation mode: Blockchain is a multi-party collaborative network architecture based on distributed systems, while federated learning requires the joint participation of multiple distributed entities to co-train the model, so blockchain can be used as the basic topological architecture of federated learning. 2) Both have the characteristics of trustworthiness: the credibility of federated learning is reflected in the fact that the privacy can be protected from disclosure in the process of data cooperation, and the credibility of the blockchain is reflected in the consensus mechanism and data verification mechanism can be used in the accounting process, so that the data cannot be tampered with and repudiated. 3) The application purpose of blockchain and federated learning is complementary to each other: federated learning aims to "create value", take advantage of the complementarity of the data of various participants, and improve the effect of the model through joint training. Whereas, blockchain is designed to "transfer value", truthfully record the contributions of all parties involved, and reward them. Therefore, the integration of blockchain and federated learning will become a development trend and the main research motivation of this paper[18−19].

At present, blockchain-based federated learning research is still in its infancy, and most of the existing studies are based on the application mode of blockchain + federated learning in combination with typical scenarios such as edge computing, Internet of Things, and Internet of Vehicles, and there is a lack of review articles that comprehensively summarize the architectural models, theoretical systems, and technical progress of the field itself[20−28]. Table 1 summarizes the differences between the main domestic and foreign review work and this paper in recent years. In general, on the basis of a comprehensive investigation in this field, this paper proposes a conceptual model of the integrated architecture of blockchain and federated learning for the first time from the perspective of using blockchain to improve federated learning, and comprehensively summarizes the key issues, research methods, current progress, application fields and future research directions in this field, which is expected to provide reference for the development of this field.

Table 1 A review of the BeFL study

The structure of this paper is as follows: Section 1 gives the conceptual model of the blockchain-based federated learning architecture and outlines its basic workflow; Section 2 elaborates on the main research issues and current progress in the five aspects of infrastructure, consensus mechanism, economic incentives, smart contracts, and privacy and security in the conceptual model. Section 3 introduces the application areas of blockchain-based federated learning architecture; Section 4 discusses existing technical bottlenecks and solutions; Section 5 outlines the way forward; Section 6 summarizes the content of this article.

1. Blockchain-based federated learning: a conceptual model

In order to systematically and comprehensively summarize the research status and progress in the field of blockchain and federated learning integration, this paper first proposes a blockchain-based federated learning architecture model. As shown in Figure 1, the BeFL architecture is divided into the infrastructure layer, the consensus mechanism layer, the economic incentive layer, the smart contract layer, the privacy and security layer, and the application domain layer from the bottom up. Among them, the infrastructure layer aims to use the decentralized (or weakly centralized) distributed system architecture of blockchain as the underlying framework of federated learning, so as to ensure system robustness and data trustworthiness. The consensus mechanism layer aims to learn from the existing blockchain consensus algorithm and improve it, so as to avoid waste of resources and ensure the credibility of federated learning results. The economic incentive layer mainly designs a reward and punishment mechanism to maximize the benefits of user nodes and motivate more nodes to join the federated learning process. The smart contract layer aims to replace the central server with smart contracts to automate the process management of federated learning, and serve as an interface and carrier for the integration of blockchain and artificial intelligence technology. The privacy security layer uses encryption and privacy computing technology to ensure the security and user privacy of the federated learning system.

Figure 1 A blockchain-based federated learning conceptual model

Figure 2 shows the basic operation process of the BeFL architecture, which includes three main links: authorization screening of participating nodes, federated modeling of the computing process, and trusted storage of computing results. In the federated node authorization screening link, each participant has their own private data, and the task publisher broadcasts the federated learning task by deploying a smart contract on the blockchain network, and uses the blockchain identity authentication mechanism to screen and authorize the participants. Verified participants will participate in the blockchain network as "miners", but a small number of participants will be selected as validators to participate in joint training under the consensus mechanism of the blockchain network.

Figure 2 shows the basic operational flow of the BeFL architecture

In the federated modeling process of the calculation process, according to the distribution characteristics of the data, it can be divided into three scenarios: horizontal federated training, vertical federated training and federated migration training. In horizontal federated training, the participants selected as the training nodes use their own local private data samples and refer to the training task broadcast by the smart contract to train the local model, encrypt the model parameters after the training is completed, and broadcast the update status of the local encryption parameters and the corresponding processing information on the P2P network. Compared with horizontal federated training, vertical federated training requires cryptographic entity alignment before training, and the alignment algorithms and rules are executed according to the code defined in the smart contract in advance, and the public key is distributed by the system at the beginning of training. In the federated migration training, the training nodes are divided into two sides, the source domain and the target domain, and the loss function is calculated together according to the secret sharing protocol in the smart contract. Subsequently, each participating node receives its locally trained cryptographic model parameters from the training node before the preset timestamp, and implements local model parameter aggregation according to the aggregation rules defined in the smart contract.

After the smart contract receives the aggregate model uploaded by the participating nodes, it broadcasts it in the blockchain network. At the same time, after the participants selected as validators obtain the aggregation model from the blockchain network, verify the model according to the predefined verification method in the smart contract, obtain the global model, quantify the contribution value of the training node, and make each participant update the global model through the P2P network before the next round of joint learning.

In the trusted storage of computing results, the federated learning results are stored in blocks, and each miner (participant) mines the block according to the blockchain consensus protocol until the random number that meets the requirements is found or the block generated by other miners is received. When a miner generates a new block, other miners verify the contents of that block (e.g. nonce numbers, smart contract changed states, transaction and rollup models, etc.). If a block is verified by a majority of miners, the block will be added to the blockchain and accepted by the entire network, while the participants will be rewarded according to the predefined reward mechanism in the smart contract, and the participants will be authorized and screened again to start the next round of training.

2. Key technologies and research progress

This section will describe the key technologies and research progress of the BeFL conceptual model (as shown in Table 2), including infrastructure, consensus mechanism design, economic incentive mechanism design, smart contract integration, and security and privacy protection.

Table 2 Current status of BeFL research

2.1 BeFL Infrastructure

The traditional federated learning model usually adopts a star network topology, in which a central server coordinates the communication round, iteratively broadcasts the current model to the participants, collects the gradient updates of local computation, and generates the next-generation model through aggregation. This centralized network topology may have problems such as single point of failure and lack of trust. First of all, once the central server fails, it will cause the entire network to be paralyzed, and when there are a large number of participants conducting joint training, the central server will pose a performance and communication bottleneck. Secondly, the premise of the centralized network topology design is that the participants are willing to trust the central server. However, even if this central server can guarantee the security and privacy of model parameter updates, participants may prefer to share parameters directly with each other due to a lack of trust in the central server. Finally, the advantage of multi-party participation in federated learning tasks is that the data stored on the private servers of various parties can be centrally aggregated, so as to improve the performance of the federated learning model. However, as the number of data owners involved in the computation increases, the participants may have the scenario of malicious attack on the system. In a situation where data has become a valuable resource, there is an urgent need to introduce new technical solutions to review the credibility of participants and motivate them to contribute data.

In recent years, the technical characteristics of blockchain technology in terms of data rights confirmation, identity authentication and automatic execution incentive mechanism can provide effective technical support for the construction of data federation. The blockchain system is usually based on the P2P network, where each node is equal and communicates and interacts with each other in a flat topology, without any centralized special nodes and hierarchies, and each node will assume the functions of network routing, verification and propagation of block data, and discovery of new nodes. P2P nodes provide some of their resource capabilities (such as processing power, storage power, or network bandwidth) directly to other network participants, without the need for centralized coordination from a central server.

Figure 3 shows the network topology diagram of the federated learning model and the BeFL architecture. Compared with the federated learning model, the network topology of the BeFL architecture is decentralized (or weakly centralized), and each user node has the same status and can join and exit freely, so the overall network has stronger scalability, service capabilities and load balancing capabilities. At the same time, the user node not only uses local data to train the local model, but also accepts the gradient parameter update of other nodes to aggregate the global model, so a single point of failure will not affect the overall system, and has higher network robustness and availability. Finally, all nodes have a relay forwarding function, which can greatly improve the anonymity of communication and protect the privacy of personal data[28].

Figure 3 The network topology of the federated learning architecture and the BeFL architecture

There are two ways to classify BeFL architecture in the existing literature. One is that the BeFL architecture can be divided into tightly coupled architecture and loosely coupled architecture based on whether the node undertakes both the federated model training task and the block consensus task. The former is characterized by the fact that nodes undertake both the training task of the federated model and the task of block consensus verification, that is, the federated learning network itself is a blockchain network [50, 63, 67]. The consensus node and the federated model training node of the latter are two groups of logically independent nodes that undertake their respective tasks [30, 37, 117−118]. The other can be divided into public chain BeFL architecture and permissioned chain BeFL architecture according to the integrated blockchain platform [29, 31, 62]. The former mainly focuses on the design of incentive mechanism to build the whole system, while the latter includes consortium chain and private chain, and mainly uses the design consensus mechanism to classify and select participants, so as to improve the security and efficiency of the whole system[32, 34, 67].

In the traditional federated learning architecture, it is difficult to ensure the quality of uploaded parameters. Different participants hold different datasets with different sizes and quality, which will lead to significant quality differences in the parameters they train. Even if no malicious participants deliberately upload low-quality parameters, this uneven quality will seriously drag down the training progress and accuracy of federated learning. Low-quality parameters may degrade system performance, resulting in unbearable convergence times. In this regard, although the introduction of blockchain technology cannot affect the data quality of the participants themselves, it can affect their behaviors and strategies for uploading data: through the integration of the BeFL architecture of the blockchain and the use of the data verification mechanism of the blockchain, the identity information of the participants and their upload parameters can be verified, so as to screen and reward and punish the participants, and encourage the participants to contribute high-quality data.

The verification process adopted in the existing literature implements the secure identity authentication mechanism from the following two stages: First, the training node deposits part of the collateral into the smart contract before participating in the joint training. Then, after the nodes participating in the joint training complete the joint training task, the training node will broadcast its public key on the blockchain and send the updated model parameters in the form of a transaction on the blockchain with a legal signature (generated by its private key). If the signature is valid, the miner confirms that the uploaded model is from a legitimate participant, and then puts the received model into the local model library. Miners will perform tasks such as model summarization and block mining, feed back the summarized model results to the training nodes, and report their verification results. If their signatures are invalid, then mark these nodes as untrustworthy and update their public keys to an invalid state, and the system will withhold their deposits in the smart contract[30].

Considering the limitations of storage space and communication bandwidth on the blockchain, although storing the model and model update data on the chain has security advantages, it also creates a huge burden on the storage capacity of blockchain nodes. Therefore, decentralized storage protocols such as Inter-planetary file system (IPFS) are usually used to store model parameters, and only hash values are transmitted and stored on the chain, rather than the entire model parameters or model gradients. The model training process is executed off-chain, and this on-chain + off-chain collaborative architecture design helps to improve the efficiency and applicability of the BeFL framework[31].

In the existing literature, researchers have proposed many new BeFL architectures in three aspects: decentralized architecture, identity authentication, and on-chain + off-chain architecture. For example, Swarm learning (SL) proposed by Warnat-Herresthal et al. [32] in 2021 is a BeFL computing architecture that integrates federated learning and blockchain, and its underlying network architecture adopts private permissioned chain technology, where each participant can only execute transactions with prior authorization, and new nodes need to be reviewed by smart contracts before they can join the joint training. After authorization, each participating node uses local private data for model training, passes model parameters through the blockchain network, and builds the model independently. In the case of using blockchain as the infrastructure, compared with the traditional federated learning architecture, SL eliminates the need for a central coordinator, and each node has the same power to aggregate the model parameters, which can ensure the credibility of the learning results. Kim et al. [33] proposed Blockchained federated learning (BlockFL), in which user nodes can send the parameters of the local model update to the nodes associated with it through the blockchain network after the user node performs model training locally. All local updates are cross-validated between nodes, and the node that obtains the accounting right is responsible for packaging all updates into the block and broadcasting them to other nodes for verification. Only after the verification is passed, the block will be added to the blockchain. This architecture design can effectively prevent the tampering of model parameter data in the process of federated learning and improve the system security. Through the distributed ledger mechanism of the blockchain, BlockFL can verify whether the parameters uploaded by each participant match the size of the local dataset it claims, so as to avoid the deliberate tampering or low-quality model parameters uploaded by the participants. However, BlockFL cannot access or verify the original local dataset of the participants, and the participants can still influence the model parameters by modifying the dataset. Therefore, BlockFL can only prevent the model parameters from being directly tampered with, but cannot prevent the participants from indirectly influencing the model parameters by modifying the dataset. This problem can be solved by other technical means, such as introducing a reputation mechanism, or allowing participants to use only verifiable datasets, such as adding digital signatures or watermarks to datasets. The BeFL architecture designed by Lu et al. [34] establishes a secure link between all terminal IoT devices through a blockchain network, and its underlying blockchain architecture only retrieves relevant data and manages the accessibility of data, rather than recording raw data. Retrieval transactions and data-sharing transactions are executed between each user node, and broadcast in the blockchain. All records are collected into blocks, which are encrypted and signed by the collecting nodes. In this way, the blockchain records all data sharing events, tracks the use of data, and ensures the auditability of the system.

2.2 Consensus mechanism design

The core of the blockchain consensus algorithm is the process of selecting one (or more) logging nodes from all consensus nodes through election, proof, alliance, random or hybrid. The advantage of the existing blockchain consensus algorithm is that any request that has already been completed will not be (or extremely difficult) to be changed, and the customer request can be credibly executed in an environment where there are Byzantine malicious nodes, so it has a high level of security[35]. In the federated learning system, how to maintain the security of the system and improve the performance of the system by trustfully screening the participating nodes and automatically executing rewards and punishments has always been an urgent problem to be solved. Therefore, it is one of the important design purposes of BeFL architecture to design the federated consensus mechanism to alleviate the security and system performance challenges faced by the federated learning system by using blockchain technology, which is mainly reflected in the following aspects.

First of all, in the underlying blockchain network architecture of BeFL, each node is an equal subject, and the aggregation function of the central coordinator is also undertaken by each training node. Therefore, the blockchain consensus mechanism ensures the consistency of the distributed ledger, and this design concept also ensures the consistency of BeFL nodes when the parameters are updated, and makes them obtain a common global model. Secondly, the consensus mechanism of the blockchain can incentivize honest nodes and punish malicious nodes. For BeFL participating nodes, there are also adversary nodes that use malicious behavior to attack the system, which requires a consensus mechanism to punish adversary nodes and allow honest nodes to gain more power. This is a good supplement to the traditional federated learning defense mechanism, which can ensure the robustness of the whole system by the mechanism design of the whole system.

In the existing literature, researchers have designed a variety of federated consensus mechanisms suitable for BeFL by drawing on the mainstream blockchain consensus algorithms. This section summarizes the research status and progress of the BeFL federal consensus mechanism from the perspectives of three types of consensus: election, proof and alliance mechanism.

1) In electoral consensus (e.g., Paxos and Raft), miner nodes elect the current round of accounting nodes by voting during each round of consensus. BeFL can use the electoral consensus algorithm to select the nodes participating in the training, maintain the robustness of the system by selecting honest nodes, and reduce the number of communication rounds due to the selection of nodes, so as to further improve the efficiency of collaborative learning by all parties[36]. In the process of joint training, it is necessary to consider the data quality, the number of computing resources, the current power, and the network conditions of the nodes, and select the nodes according to certain weights[37]. According to the candidate selection algorithm with decreasing weight, the BeFL network will select the appropriate device to join the training in each joint training round. Even if some of the participating nodes exit, the BeFL framework can still effectively help the remaining nodes continue training to get a better model[38].

2) In proof-of-consensus algorithms (e.g., Proof of work/PoW and Proof of stake/PoS), miner nodes must prove that they have a specific ability in each round, and the proof method is usually to competitively complete a task that is difficult to solve but easy to verify, and the miner node that wins the competition will get the right to book. Inspired by this, a similar consensus algorithm is designed in the existing BeFL research to ensure the robustness of the system. For example, Chen et al. [39] designed a dedicated proof-of-stake PoS consensus mechanism. This mechanism protects correct and legitimate local model updates by rewarding stake to honest devices more frequently, thereby increasing the chances of honest nodes dominating blocks. This design promotes the honest device to participate in joint training more frequently, so as to obtain a better global model. The simulation results show that if the MNIST handwriting dataset is used for joint training, the accuracy can reach 87% in the environment with 15% of malicious devices. Similarly, Kang et al. [40] proposed a PoV (Proof of verifying) consensus algorithm, which can defend against poisoning attacks and ensure reliable federated edge learning by removing low-quality model updates and managing qualified model updates in a decentralized and secure manner. Furthermore, the PoC (Proof of contribution) consensus algorithm designed by Zhu Jianming et al. [20] comprehensively considers the contribution of nodes to the whole BeFL from three aspects: node online time, local model quality and data contribution. Compared with the traditional proof-of-proof consensus algorithm of the blockchain system, it saves the computing resources for solving the meaningless cryptographic problem, prevents the selfish behavior of the node, and effectively ensures the fairness of the whole BeFL reward distribution. In addition, an important problem in federated learning is that with the rapid increase of data volume, it is necessary to filter out relevant data from many irrelevant data. In order to solve this problem, Doku et al. [41] used blockchain technology to design and propose a PoCI (Proof of common interest) consensus mechanism to solve the problem of scarcity of related data. In the design of this consensus mechanism, the whole network is composed of different interest groups. The goal of each interest group is to train its own global model, and the potential node that wants to join the interest group will execute a computation task, and the interest group will judge whether the result of the computation task is consistent with its own interests, so as to decide whether to absorb the potential node into its own group. Therefore, this mechanism can filter the data of the training node, which is helpful to alleviate the problem of data scarcity.

3) In consortium consensus (e.g., Delegated proof of stake/DPoS), consensus nodes determine a group of representative nodes based on a specific method, and then the representative nodes obtain the accounting rights in turn or election. In the BeFL network architecture, with the disappearance of the central server, the pressure of computation and network transmission will be transferred to the training nodes, especially when all nodes have to handle the consensus task, and the computational overhead of each round is very large. Therefore, inspired by the idea of consortium consensus, a committee-based consensus mechanism is designed, which can not only improve the training efficiency, but also optimize the training process of machine learning[42-43]. The consensus mechanism determines a group of honest nodes, forms a committee responsible for verifying the local gradient and generates blocks, and the remaining nodes perform local training and send local updates to the committee. The committee's local data will then validate the updates as a validation set, and a score will be given to them, and only qualified updates will be packaged onto the blockchain. At the beginning of the next round, the system will select a new committee based on the scores from the previous round. Using the datasets and evaluation functions published by the organizers, users can compete for the first or best training model, so as to maximize the evaluation function. It can be seen that BeFL selects a committee through the blockchain to aggregate the model and record verifiable proofs in the blockchain, thus improving the verifiability.

Different from the above three ways of using blockchain to design federated consensus mechanism to improve the performance of federated learning, researchers are also committed to designing a new BeFL architecture and using federated learning to improve the consensus algorithm of blockchain system. For example, the commonly used Raft consensus protocol is a simple, leader-based consensus algorithm in which only the leader can process all client requests, copy logs, and transmit them to nodes. As a result, many private blockchains use Raft as their underlying consensus algorithm. However, if the leader has network instability issues such as packet loss, the use of the Raft algorithm will increase the likelihood of network splitting, and more than half of the nodes will lose control of the current leader. When a network split occurs, Raft will begin a leadership election period. During this time, the network will stop processing requests from clients. This leads to an increase in transaction latency and a decrease in Transaction per second (TPS) throughput. Therefore, combining federated learning to analyze and predict the network log data of each node can not only identify the factors affecting network stability and select a new leader to minimize network splitting, but also improve the performance of the Raft algorithm by shortening the block decision time and flexibly selecting a node that can produce better performance according to the state of the network environment[44]. For another example, the PoW consensus algorithm adopted by Bitcoin determines the bookkeeping power and reward through competition between nodes (miners) to solve the hard cryptographic puzzle through intensive computation, which is a very computationally resource-intensive process. The energy waste problem of PoW deviates from the current trend of sustainability and environmental friendliness in the development of technology, thus diluting its value and hindering its further application. BeFL can replace the cryptographic puzzle in PoW with federated learning tasks, and transform meaningless hash proofs into actual tasks to complete the training process of federated learning models, thus expanding the application scope of blockchain[45].

2.3 Design of economic incentive mechanism

The effect of federated learning depends on the quality of node local model updates. However, nodes may be reluctant to participate in joint training and share their model updates without sufficient incentives. Unlike cloud-based distributed machine learning, the participating nodes in the BeFL architecture are independent, and the data owner can determine when, where, and how to participate in federated learning. Furthermore, participation in federated learning will result in the consumption of computing resources and network bandwidth, and only sufficient rewards can incentivize users to endure these costs. Therefore, rewards can be used to influence the user's decision in some way. Through different incentive mechanisms, users will execute different training strategies, which will affect the final machine learning model performance.

Blockchain is a typical economic system, and its participants can obtain economic incentives by contributing computing power. Therefore, the key to the design of the BeFL incentive mechanism is to reasonably evaluate the contribution made by each node, and attract and retain more users to participate in the joint training. The algorithm described in this section assumes that the participating nodes have datasets of different quality, so as to investigate whether different incentive mechanisms can reward the nodes participating in the training fairly and justly, and attract more nodes to join the joint training. In the existing literature, many researchers are committed to designing the BeFL incentive mechanism according to the user's contribution, and the incentive mechanism can be designed from three aspects: user data quality, user behavior and user reputation.

2.3.1 Incentive mechanism based on user data quality

In terms of data quality verification of participants, it can be achieved by executing specialized algorithms in federated learning, and the use of blockchain will add a certain amount of complexity, and may bring some performance and resource overhead. However, the main reason for data quality verification by combining blockchain and federated learning lies in the credibility and security requirements in the process of data sharing and model aggregation. Federated learning involves multiple participants cooperating to train models, and these participants are often distributed in different geographical locations or organizations, which presents challenges in trust and data security. The decentralization, immutability, and transparency of blockchain can enhance the credibility and security of data exchange and model aggregation processes. In terms of data quality verification, blockchain can ensure that the history of data exchange is publicly verified, preventing data tampering or malicious behavior. At the same time, blockchain can provide an incentive mechanism, encouraging participants to provide high-quality model updates, increasing the enthusiasm and motivation of participants.

SV (Shapley value) is a typical method for assessing the quality and value of data. The SV method originated from game theory, and is widely used in the rational distribution of benefits in economic activities. The calculation of SV needs to consider the average marginal utility of data points in different subsets, that is, the impact of adding or removing data points on the performance of the model. The higher the SV, the greater the value of the data points to the model[47]. Liu et al. [46] proposed a BeFL framework called FedCoin, in which user nodes compute SV and create new blocks based on the PoSap (Proof of shapley) consensus protocol. Based on the SV calculated by the user's nodes, FedCoin will implement an incentive income distribution scheme with repudiation and tamper-proof characteristics. Experiments show that FedCoin can correctly determine the contribution of federated learning users to the training of global models, and reach the upper limit of computing resources required to complete PoSap consensus. Similarly, Ma et al. [47] are blockchain-based to measure data owners' SV-based contributions with configurable resolution without sacrificing their privacy, which solves the challenge of transparent contribution evaluation by different data owners in a federated learning environment across data silos. However, SV also has some drawbacks, which require the training and evaluation of machine learning models with different combinations of training datasets when calculating the contribution index of different nodes, which will consume more time and resources. When the node produces adversarial behavior, SV cannot accurately measure the data quality.

In addition, evaluation criteria can be set for training models. For example, Mendis et al. [31] investigated the distributed machine learning and federated learning mechanisms for Ethereum blockchain rewards. Participants use local data to train a global model and upload model parameters via IPFS. If the uploaded model evaluation value exceeds a predefined minimum acceptable applicability threshold, participants will be rewarded in Ethereum. Martinez et al. [48] proposed a class-like sampling verification error scheme based on the smart contract verifying and rewarding only valuable upload updates, and examined and evaluated two verification criteria, error trend and error threshold, to determine the quality or usefulness of the local data used to train the model.

2.3.2 Incentives based on user behavior

Analyzing user motivation and motivating them to choose the right behavior is also an important topic for researchers. For example, in the BeFL architecture, the training node may not perform local learning, but directly copy the uploaded model parameters from other clients to save its computing resources, so that the lazy node can invest more computing resources to mine blocks and thus obtain more block rewards[49]. This behavior is unfair to other nodes that use computing resources for model training. Therefore, BeFL usually formulates strict reward policies based on the incentive compatibility of competition theory in game theory, so that any rational participating node can follow the protocol and maximize its own interests by formulating strict reward policies for the number of nodes participating in the training, the reward distribution method, and the node contribution[50].

Traditional federated learning generally adopts the crowdsourcing model to design the reward mechanism. For example, in the federated learning framework designed by Pandey et al. [51], the crowdsourcing model is used to facilitate the participation of multiple devices in federated learning under the consideration of computational and communication cost-effectiveness. Based on Stackelberg game theory, this paper studies the interaction patterns between participating clients and application platforms, so as to maximize the benefits of participating users and construct high-quality learning models. This approach can be well combined with blockchain to design decentralized self-organizing incentive mechanisms [52-53]. For example, the BeFL framework 2CP (Crowdsource protocol and consortium protocol) designed by Cai et al. [54] supports the crowdsourcing model, in which user nodes propose models for training and use their own data to evaluate their contributions. Users will be rewarded based on their relative contributions, and users with larger or higher quality datasets will receive a higher share of the rewards in the final model.

2.3.3 Incentive mechanism based on user reputation

Reputation is an important indicator for selecting user nodes in the process of federated learning. Users with higher credibility are more likely to bring high-quality and reliable training to federated learning tasks. At the end of each training task, the user's reputation is updated based on the user's behavior, and then the reputation record is taken into account when the user is selected in the next training. When a user contributes correct and useful model parameters, its reputation increases, while when malicious model parameters are uploaded, its reputation decreases. There are many benefits to doing so, such as reducing the time cost when blockchain nodes are voting, where reputable nodes can avoid complex majority voting for verification. When dealing with forks, if a node receives two or more satisfactory models from different nodes at the same time, a natural solution is to choose the node with the highest reputation.

When designing the BeFL reputation scoring mechanism, it is necessary to ensure the authenticity and objectivity of the reputation score. Specifically, the edge nodes, fog nodes and cloud servers participating in federated learning can be graded according to their respective data quality and participation degree, and score each other. The reputation of each participant in federated learning is then recorded through smart contracts[55]. In order to make the reputation score more objective, each participant needs to be scored from multiple dimensions. In the BeFL architecture designed by Kang et al. [56], the task publisher uses the RONI (Reject on negative influence) scheme for the independent and equally distributed scheme and the FoolsGold scheme for the non-independent and equally distributed scheme to detect attackers and unreliable clients, and update the user reputation score based on the detection results. The reputation score of each user will be generated by combining the different scores given by all task publishers and combining the weights[57]. Thus, relying on the reputation scoring mechanism, the honest behavior of users and high-quality data will benefit themselves and promote the healthy development of the entire system. The BFL framework proposed by Qi et al. [58] designs a reputation evaluation mechanism based on model quality and a reward distribution algorithm based on reputation-weighted contribution to incentivize data owners to provide high-quality data. The algorithm takes into account the data volume, reputation value and unit resource consumption of the data owner, and solves the optimal data contribution strategy. In order to analyze the behavior strategy of the data owner, a non-cooperative game model is established, which proves that the game model has a unique Nash equilibrium, that is, each data owner has no motivation to deviate from its optimal strategy, and the validity, safety and reliability of BFL are verified by simulation experiments and theoretical analysis.

2.4 Smart contract integration for BeFL

A smart contract is a computer program on a blockchain that operates in a decentralized and autonomous manner that can send, receive, and store information, and respond to input information based on predefined execution logic. Once deployed on-chain, smart contracts are undeniable and tamper-proof. Smart contracts are typically written in a specialized programming language (such as the Solidity language on Ethereum) and executed through a blockchain-based virtualized environment. This virtualized environment provides a common platform for creating, testing, and deploying smart contract-driven decentralized applications (DApps). Based on smart contracts, we can design diversified scheduling rules such as federated learning task release, local parameter verification, global model aggregation, global model release, node contribution evaluation, and incentive mechanism implementation, which is helpful to realize efficient adaptive optimal scheduling of decentralized federated learning process[59−60].

The integration research of smart contracts and federated learning usually follows the mainstream smart contract programming languages (such as Ethereum's Solidity language). For example, Li et al. [43] use the traditional Solidity smart contract to manage the federated learning process, and use the precompiled smart contract to design the corresponding functional function module (the contract bytecode data uploaded from the chain needs to be transformed when performing federated learning training locally), and listen to model parameters, preset model aggregation rules, and broadcast global model parameters. In general, in the BeFL architecture, the functions of the central server are instead implemented by decentralizedly executed smart contracts to automate the coordination of workflows among distributed nodes. These functions can be roughly divided into three main parts: participant management, federated modeling process management, and incentive mechanism management.

1) In the participant management part, for each participant participating in federated training, there is a smart contract with the same preset algorithm corresponding to it, and it is executed separately in each participant's computing environment. The smart contract preliminarily screens the participating nodes through the decentralized digital identity management system. Authenticated nodes initiate modeling tasks, and smart contracts start data-driven learning tasks based on the details of the corresponding events, such as the properties of the input data, the expected output of the task, and the incentives.

2) In the federated modeling process management part, the participating nodes use local private data off-chain, train the model for published tasks, and use a distributed storage system (such as IPFS) to store the parameters of the model. By importing the returned hash value into the DApp, the training node can broadcast the implemented computational model. After receiving the computational model, the validator starts the evaluation and reports the results of its evaluation. The smart contract obtains the evaluation results of each validator, and then checks them against the verification criteria pre-set by the task publisher in the smart contract. If the number of evaluation results that meet the criteria exceeds a certain threshold, the system will automatically distribute financial incentives to the accounts of the relevant training nodes, and all validators will receive financial incentives.

3) In the incentive mechanism management part, due to the possibility of incorrect or even malicious node behavior during the training process, the training and validator nodes are usually required to stake a certain amount of tokens to the smart contract at the beginning of each round of training. If the nodes correctly participate in the joint training process, then after the training ends, the training and validator nodes will receive the corresponding service remuneration and staked tokens according to the commitments agreed in the smart contract. If malicious or anomalous behavior is detected, the node will not be paid, and its staked tokens will be distributed to all other participating nodes. Combined with the immutability of the underlying blockchain, smart contracts can supervise, regulate and trace the behavior of all parties involved, and automatically implement the reward and punishment incentive mechanism.

It can be seen that there are many advantages to using smart contracts to manage the federated learning process: firstly, the global model copy and related computational state are maintained in the smart contract, the model selection and aggregation are carried out in a decentralized manner, and the participating nodes can determine their own choices, which helps to establish trust between nodes. Participating nodes can use a copy of the global model to autonomously perform aggregation steps in each turn, and update the global model independently, thus promoting the development of global computing. With the support of the underlying consensus protocol, smart contracts help to ensure the autonomy, fairness, and impartiality of the federated learning process. Secondly, smart contracts can coordinate multiple federated learning task rounds and model parameter aggregation from different user device sets at the same time. Compared with the federated learning model based on the central server, federated learning based on smart contracts can significantly reduce the setup and operating costs, thereby lowering the threshold of federated learning training and helping to attract more users to join the federated training. Finally, smart contracts can supervise and normalize the behavior of all parties involved, and automatically implement incentives and punishments[61].

In addition, various types of smart contracts running on the blockchain can be regarded as intelligent agents of users. At this stage, smart contracts can only execute predefined trigger actions according to preset rules, which is not yet intelligent. The integration of blockchain and federated learning will help machine learning, deep learning, reinforcement learning and other artificial intelligence technologies to be integrated into the BeFL architecture in the form of smart contracts, forming intelligent components for different tasks and scenarios, so as to promote smart contracts from basic capabilities such as task selection, prioritization and goal-oriented behavior, to intelligent agents with high-level capabilities such as perception, reasoning, learning and autonomous decision-making. Decentralized autonomous organizations (DAOs) are formed with certain social skills, and their integration and evolution process is shown in Figure 4.

Figure 4 Integration and evolution of smart contracts and artificial intelligence

In terms of the integration of smart contracts and artificial intelligence, researchers have proposed to use the machine learning algorithm XGBoost to detect Ponzi schemes on the blockchain [62], use smart contracts embedded with support vector machine algorithms to process the unbalanced traction and braking data generated during train operation [63], and use smart contracts based on ensemble learning models to optimize the classification tasks of automatic modulation in radio [64]. In addition, deep learning algorithms can also be integrated with smart contracts to use autoencoders to detect anomalies in network intrusion [65–66], but further research is still needed for the integration of deeper parameter models. Combined with reinforcement learning, it can be used to screen BeFL nodes with good computing power and good communication, and can optimize the computing and communication resources occupied by the system, so as to improve the operation efficiency of the system[67−69].

Furthermore, after combining artificial intelligence technology, BeFL can be regarded as a distributed intelligent system composed of many intelligent agents, so that complex tasks can be divided into interrelated subtasks from top to bottom, and these subtasks can be solved by dividing and conquering by intelligent agents, and the business of the actual system can be solved from the bottom up through high-level emergence[70]. For example, Zhang et al. [71] designed a BeFL framework based on intelligent agents (SABlockFL), in which intelligent agents act as consensus nodes and participating nodes of federated learning tasks in the blockchain network at the same time, and use machine learning task datasets (such as MNIST and CIFAR) for joint training, and the experimental results prove the effectiveness of the framework.

From the perspective of social intelligence or swarm intelligence, federated learning based on smart contracts will have a profound impact on community-centered organizational forms. Micro-organizations in the same community can use smart contracts on the blockchain to achieve self-organizing joint training[72]. In this way, every organization in the community can collaborate to train public tasks to help solve challenges common to the entire community while preserving data privacy. Furthermore, by encoding the core legal provisions, business logic, and intent agreements into smart contracts, it is possible to realize the "law-based chain" for specific business scenarios and regulatory needs, and build a compliant, trusted, data-sharing, and regulator-friendly BeFL multi-agent system, which will gradually evolve into a federated ecology formed by various DAOs[73]. This will promote the development of distributed artificial intelligence technology, and promote the transformation of traditional business models and social production relations.

2.5 Security and Privacy

The main advantage of federated learning is that each round of model training is completed locally, and a copy of the global model is retained. During each round of training, the data is stored on the end device rather than on a cloud server, which can greatly improve privacy and help meet the requirements of regulations such as GDPR. However, there are also security and privacy challenges in the training process of federated learning. The main attack methods in federated learning security include poisoning attacks, backdoor attacks, Byzantine attacks, free-rider attacks, and sybil attacks, and the main attack methods encountered in privacy are inference attacks and adversarial attacks[74−76].

The BeFL architecture deploys federated learning on the blockchain network, which can better protect data privacy and security by taking advantage of the encryption, immutability, and decentralization of the blockchain. For example, data and model parameters are encrypted and written to the blockchain when they are updated, reducing the risk of being stolen by intermediate third parties. By storing the model parameters or updates of federated learning on the blockchain, it can be ensured that the training history of the model is traceable and transparent. This helps to detect and prevent malicious attacks, as all operations are recorded and any unauthorized access or tampering is immediately detected; In addition, blockchain technology can also be used for authentication and access control to further improve the security of federated learning systems[81−82].

After combining blockchain technology, although the BeFL architecture uses blockchain technology to screen federated learning participating nodes, automate the execution incentive mechanism, and adopt trusted block storage, etc., to help the federated learning system alleviate the security and privacy problems to a certain extent, it still cannot completely solve the problem. The main idea of the related research is to investigate the robustness of the BeFL architecture by assuming that no more than half of the participating nodes are selected as dishonest nodes to simulate the attack on the system. In this section, we will first discuss the privacy protection mechanism of the BeFL architecture, and then introduce the challenges that are currently focused on solving in the research of the BeFL architecture, such as inference attacks and poisoning attacks.

2.5.1 Privacy Protection

In the existing literature, a variety of privacy protection mechanisms corresponding to different business scenarios are proposed. For example, homomorphic encryption, secure multi-party computation and differential privacy are the most widely used privacy protection technologies: homomorphic encryption aims to perform computational operations on the ciphertext and obtain the encryption result, and the encryption result is the same as that of the plaintext after decryption. Secure multi-party computation aims to enable multiple participants to co-compute a function without exposing their own private input data. Differential privacy aims to hide personal information from client model updates by adding a small amount of noise. Based on these privacy-preserving technologies, researchers have proposed many BeFL design frameworks [77−81].

It is worth noting that in the design of the BeFL framework, it is necessary to consider reducing the heavy computational and communication overhead, for example, when using differential privacy, higher privacy requirements mean that more noise needs to be added to the query results, if users use differential privacy to protect their parameters, then they need to send more bits to ensure the accuracy of the aggregation results, and different participants need to communicate with each other to coordinate the process of adding noise, so higher privacy requirements mean higher communication costs[82]. Secure multi-party computation protocols require all parties to generate a secret share of their private data and exchange it with other parties, thus requiring multiple data transmissions, which inevitably leads to higher communication overhead. The computation and transmission process of homomorphic encryption are carried out in the ciphertext state, so the computation time is longer and the transmission is slower. It can be seen that communication overhead, computational time, and deployment environment are the problems that BeFL must consider when designing the privacy protection mechanism. In order not to reduce the prediction accuracy in the case of differential privacy, the FPPDL (Fair and privacy-preserving deep learning) proposed by Lyu et al. [83] reserves a part of the insensitive data for global sharing at the beginning of training, and generates the training data by using a Generativeadversarial network (GAN). Deep learning is realized based on the synthesized data samples. In the first phase, FPPDL uses Differentially Private GAN (Differentially Private GAN) to publish differentially private local samples for mutual evaluation in the initial benchmarking phase. In the second stage of parameter uploading, the homomorphic encryption algorithm is improved with the help of stream cipher to reduce the communication overhead. The encryption scheme of FPPDL makes it effective for high-dimensional data encryption, which partially solves the problem that large models with high-dimensional parameter vectors in federated learning are vulnerable to privacy and security attacks.

2.5.2 Inference Attacks

During federated learning training, gradient exchange may leak information related to participants' training data to adversarial participants, which is called an inference attack. Therefore, observations of model updates can be used to reason about private information. The current research focuses on inference attacks on elements such as membership, attributes, and models.

A member inference attack is given an exact data point, which the attacker aims to determine whether or not to use to train a model. For example, an attacker can reason whether or not to use specific patient data to train a disease-related classifier. In the BeFL framework designed by Liu et al. [79], it is proposed to add Gaussian noise to the model updates uploaded by the participants. Numerical results show that the framework can effectively prevent member inference attacks, thereby improving the security of federated learning in 5G networks.

A feature inference attack is when an attacker infers the properties of the training data of other participants. Attackers can use multi-task learning to trick federated learning models to learn to do a better separation of data with and without attributes to extract more information. For example, when there is an attribute leak in a participant's model update, an adversary will be able to identify a group of participants with a specific attribute. In order to solve this situation, Shen et al. [84] designed a BeFL framework based on blockchain technology, and the experimental results show that the framework can keep the main tasks in federated learning unaffected.

The model inversion attack aims to extract the training data or the feature vectors of the training data from the model, and the attacker simulates a model similar to the original model by "query-responding" the data. Fang et al. [85] demonstrated that blockchain technology can provide a reliable solution for federated learning in the face of model inversion attacks.

2.5.3 Poisoning attacks

Poisoning attack is another common security attack in federated learning. According to the attack mode, it can be divided into data poisoning attack and model poisoning attack. The former refers to the attacker poisoning the data by directly injecting the poisoned data into the target device or injecting the poisoned data through other devices, thereby reducing the accuracy of the overall model. The latter refers to an attacker's attempt to poison the local model instead of the local data, such as poisoning the local model update before sending it to the server, or inserting a hidden backdoor into the global model to introduce errors that affect the accuracy of the global model.

Blockchain technology can effectively mitigate poisoning attacks. Before the start of federated learning, the participant verification mechanism can screen out external attackers. In each round of training, the review of nodes and upload parameters will cause the system to reject the data and models that have an incorrect impact on the global model, and the underlying consensus protocol ensures that the whole system can still run correctly even under the influence of a small number of errors. In addition, the incentive mechanism enables each node to choose the right way to participate based on the assumption of maximizing its own interests, and the openness and transparency of the smart contract makes each node's wrong behavior will be traced back to responsibility. All these measures help to ensure the robustness of the BeFL architecture in the event of poisoning attacks[86].

In the existing literature, Biscotti, the BeFL framework proposed by Shayan et al. [87], provides data privacy and security by using differential privacy noise and Shamir secret sharing scheme, which can provide secure private multi-party machine learning in a highly distributed environment. Experiments show that Biscotti can resist data poisoning attacks, and when 30% of opponents try to poison the model, Biscotti can protect the privacy of parameter updates of training nodes and the performance of the global model on a large scale. BlockFlow is a privacy-preserving solution for federated learning, which considers the use of Laplace noise to protect differential privacy to avoid leaking client information to other clients. In model poisoning attacks, BlockFlow can detect proxies that provide bad models and resist attacks by no more than 50% of malicious proxies without sharing personal privacy information[88].

3. Application scenarios of BeFL

At present, the BeFL architecture has been studied and applied in many fields, and is committed to solving key problems such as security and privacy protection in data sharing and multi-party collaboration scenarios, and improving the value of data elements in sharing and circulation. This section will summarize the application status and progress of BeFL in typical fields such as federal cloud computing, healthcare, Internet of Vehicles, smart city, and mobile network. The common problems faced by these fields are multi-party sharing and collaboration based on incomplete data and their privacy protection in weak trust environment. On the one hand, it may be difficult for institutions to fully share necessary data due to a lack of trust. On the other hand, market competition or industry privacy constraints may also make organizations reluctant to share data. Therefore, breaking through the dilemma of data silos and promoting data sharing and cross-chain circulation between institutions while protecting data privacy are the widely used requirements and important application goals of BeFL architecture and related technologies in these fields.

It should be considered that in a common cross-device scenario, the scale of the participants is very large, which may reach millions or even tens of millions. This situation usually faces a huge overhead of information exchange: the voting of the ledger node in each round of consensus process, and the node undertakes both the consensus task and the federated aggregation task, which can lead to performance bottlenecks and scalability problems (especially for tightly coupled architectures, but not common for loosely coupled architectures where consensus nodes and federated nodes are independent of each other). Therefore, in order to improve the efficiency of network bandwidth, computing resources and consensus algorithms, the existing research uses the consensus mechanism based on election class to select the participating nodes on the one hand, and uses the technology based on gradient compression or differential privacy to reduce the amount of uploaded data and protect data privacy on the other hand. For cross-silo scenarios with a relatively limited number of participants, it is also necessary to use blockchain to increase the trust between participants in a weak trust environment (such as cross-regional and cross-border application scenarios), and implement public traceability and incentive mechanisms in combination with blockchain to maintain the reliability and fairness of the system.

Figure 5 depicts the general pattern of the BeFL architecture in the application scenario: First, in the data access and preprocessing stage, all parties need to load the access data into the trusted computing environment, and perform heterogeneous data fusion and other preprocessing operations on multi-source data from different user terminals (such as the cloud of federated cloud computing, fog nodes, and data uploaded by the edge), and provide private set intersection (PSI) privacy sample alignment, differential privacy, and Algorithm module library for homomorphic encryption, secure multi-party computation, and federated learning model.

Figure 5 Scenarios for the BeFL architecture

Secondly, after the data preprocessing is completed, the smart contract state change will be triggered, and the smart contract will dispatch all parties to perform federated computation, and after the federated computation is completed, the winning node will package the data block into the blockchain and automatically execute the reward distribution, the detailed workflow and smart contract function module design refer to Figure 2 and Section 2.4.

It should be noted that there are two main architectures of BeFL: loosely coupled and tightly coupled. The loose coupling architecture is characterized by the fact that the federated computing nodes and the blockchain consensus nodes are independent of each other, and after the nodes of the federated learning system complete the training, the smart contract will trigger the smart contract state change, and the smart contract will dispatch the blockchain system to accept the model parameters in a unified format broadcast from the federated learning nodes. The loose coupling architecture is suitable for dynamic scenarios such as the Internet of Vehicles and mobile networks, and its advantage is that it can not affect the existing internal computing logic and underlying code of the two systems, and only needs to formulate the data interaction form that can be executed by both parties in advance and agree on the execution order to achieve the synchronization of the computing process between the two. The disadvantage is that the cost of computing and communication is large, and the communication transmission will be unstable due to the computational environment, which will affect the stability of the system. Different from the loosely coupled architecture, the participating nodes in the tightly coupled architecture are used as both the training nodes of the federated learning system and the consensus nodes in the blockchain system to cross-verify the local model. The advantage of this architecture design is that the data is processed within the node, and the transmitted data can be guaranteed to be desensitized data, which can effectively alleviate the additional computing resources and communication overhead generated when the data is converted between the federated learning system and the blockchain system, as well as the risk of data leakage when the data is transmitted between the systems, and is suitable for the application scenarios with disaster recovery and backup requirements for data in the federated cloud computing scenario and the strong sensitivity to data privacy in the medical and health scenario. Similarly, the tightly coupled architecture is difficult to be compatible with other federated learning systems or blockchain systems, and the cost of technology upgrade is large, and it is easy to form a single BeFL ecological closed loop.

Finally, the API (Application program interface) service of BeFL architecture directly serves various application scenarios and directly interacts with the participating nodes in the form of DApps. BeFL usually provides services such as federated modeling task submission, task approval, calculation result query, state analysis, and data statistics, etc., and builds a fair and credible federated computing environment in different application scenarios through these API services. In the following article, we will elaborate on how the BeFL architecture can alleviate the challenges of data sharing and privacy protection in different application scenarios.

3.1 Federated Cloud Computing

Federated cloud is an important development trend of cloud computing, and its basic idea is to enhance cloud devices through the connection and interaction between edge computing, fog computing and terminal cloud computing. For enterprises, the federated cloud is a non-synchronous off-site data storage model that greatly improves the enterprise's data disaster recovery capabilities, protects off-site data assets, and fog computing and edge computing can be used to alleviate the delay caused by factors such as computing resource limitations and network signals when data is uploaded from terminal devices to cloud servers. However, for fog computing and edge computing, it is still necessary to upload local data to fog computing and edge computing platforms, which inevitably leads to data privacy and data security issues. Therefore, under the increasingly strict data privacy policy, the integration of blockchain technology and federated cloud computing model is becoming more and more obvious, and the combination of the two can effectively support the remote secure storage of digital assets, which will surely gain more widespread application [89−94].

First of all, as the infrastructure of cloud computing and the Internet of Things, fog nodes can solve network congestion and latency and achieve local autonomy, but at the same time, they still face privacy problems, poisoning attacks and consequent inefficiencies. The BeFL framework proposed by Qu et al. [95] uses a blockchain with a distributed privacy protocol to replace the central authority, and only saves the pointers of global model updates on the blockchain, while storing the relevant raw data in the off-chain distributed hash table, so as to protect privacy, reduce poisoning attacks, and improve the efficiency of model operation.

Secondly, for the edge nodes in the federated cloud, Majeed et al. [96] designed a BeFL framework for edge networks, so that federated learning can be applied to 5G-enabled intelligent IoT applications. In this framework, a blockchain network of edge devices stores local model updates from user devices in separate channel-specific blocks in the form of blocks, and global model updates are stored in a Merkle Patricia tree on the channel-specific ledger.

3.2 Healthcare

In the field of healthcare, next-generation artificial intelligence technologies such as machine learning continue to promote the implementation of applications such as intelligent image-assisted diagnosis in medical institutions, and when using machine learning to train data, the primary concern of medical institutions is that the training data needs to be kept locally and avoid inferring sensitive information from the trained model, which makes it difficult to share these sensitive data among medical institutions. At the same time, insufficient data sources and insufficient labels will lead to unsatisfactory performance of machine learning models, which has become a bottleneck in the development of smart medical and health fields. In this regard, blockchain and federated learning can be used to ensure data security while realizing medical and health data sharing[97−106]. At the same time, the introduction of blockchain technology will enable the untrusted medical data federation to effectively build a complete and secure process that can be traced and accountable. For example, El et al. [107] proposed a federated learning algorithm based on smart contracts to implement a coordination server to ensure transparency and permission for medical data sharing, and apply it to diabetes prediction and decision support. The BeFL framework proposed by Qayyum et al. [108] can collect data from different sources collaboratively while protecting patients' data privacy, and uses a deep learning model to detect new coronavirus pneumonia images from lung image scans, which improves the level of intelligent diagnosis and treatment in hospitals.

3.3 Internet of Vehicles

The Internet of Vehicles has the characteristics of collaborative environmental data perception, real-time data calculation and processing. Intelligent vehicle data in the Internet of Vehicles environment refers to various data generated during the operation of intelligent networked vehicles, including vehicle status, driving trajectory, internal and external environment, driving behavior, user preferences, etc. This data can be used to provide services such as vehicle control, infotainment, safety protection, remote diagnostics, etc., as well as to optimize applications such as traffic management, road condition monitoring, and travel planning. Big data and artificial intelligence technology have made data sharing between smart cars a development trend. However, data involving personal identity, location, habits, and other information, as well as data that affects vehicle functions and safety performance, need to be protected and securely encrypted [109]. Therefore, how to ensure the security and privacy of data in the process of data sharing is a challenging problem at present. With the increase in the number of vehicles, traditional AI-based algorithms are becoming more and more difficult to cope with large-scale and distributed vehicle networks. The main reason is that traditional AI solutions require systems to accurately predict vehicle waiting times, and car owners may be reluctant to share personal data directly due to communication and privacy limitations. As an emerging computing paradigm, federated learning provides a feasible solution for data sharing between vehicles while protecting data privacy. However, by analyzing the parameters updated by the local model uploaded from the data provider, private information can still be leaked, and when some vehicle nodes perform malicious behaviors, it will affect the performance of the entire IoV system. To this end, Wang et al. [110] integrated blockchain, differential privacy, and mobile edge computing technologies into the federated learning framework to ensure the security and effectiveness of the Internet of Vehicles and promote data sharing between vehicles. Specifically, the designed system uses smart contracts to schedule the joint training process: firstly, the smart contract authenticates the identity of the participating nodes, and filters the transmitted model parameters to ensure the reliability of the model. Subsequently, the smart contract quantifies the model parameters uploaded by the participating nodes into the total weight according to the preset evaluation criteria, and uses them as the standard for the profit distribution of federated tasks. Posner et al. [111] optimized the federated learning system based on blockchain for analyzing and processing vehicle communication networks, which updated and verified the parameters of the on-board machine learning model among each vehicle in a decentralized manner, and realized the on-board machine learning by using the consensus mechanism of the blockchain without any data center to coordinate. Chai et al. [112] proposed to use a hierarchical blockchain framework and a hierarchical federated learning algorithm to share machine learning model parameters, which can enable vehicles to obtain more environmental data, especially for large-scale in-vehicle networks.

3.4 Smart Cities

Data is an important factor in the construction of smart cities. In the process of applying artificial intelligence technology to build smart cities, there are problems such as low data utilization rate and low model accuracy, which are due to the fact that urban data is composed of many data islands, such as traffic data, industrial data and resident data. How to solve the problem of data sharing between data islands and provide more secure and trustworthy data services is a problem that must be faced in the construction of smart cities. Federated learning allows different participants to train a model together without sharing raw data. This solves sensitive data privacy concerns. However, there are scenarios where cross-organization or cross-city data sharing is required, and blockchain can provide a secure and trusted way to exchange data between organizations. Blockchain technology can record the history of data exchange, ensuring transparency in the source and use of data, thereby facilitating data cooperation between more organizations. Federated learning involves multiple participants in the process of model update, but in this process, it is possible that some malicious participants may try to tamper with the model or provide incorrect updates. In this case, the blockchain can be used to record the update history of the model, guaranteeing that the actions of the participants are publicly verified, thus increasing the trust of the entire process. Their combination in smart cities can provide a more secure, transparent, and trusted framework for cross-organizational data collaboration. Therefore, BeFL-based data security sharing mechanism can be effectively constructed in smart cities, and it is helpful to solve the problems of increasing data complexity, changing environment, and traditional security attacks in smart cities[113−116]. For example, based on BeFL, an artificial intelligence algorithm can be used to intelligently classify network traffic in an IIoT system to identify network anomalies caused by adversary attacks[117].

3.5 Mobile Networks

Emerging technologies such as digital twins and sixth-generation networks (6G) have accelerated the realization of edge intelligence in the Industrial Internet of Things. The integration of digital twins and 6G bridges the gap between physical systems and digital space, and enables powerful instant wireless connectivity. With the increasing concern about data privacy, federated learning has been seen as an important solution for deploying distributed data processing and learning in wireless networks. However, unreliable communication channels, limited resources, and lack of trust between users hinder the effective application of federated learning in mobile networks. Combining blockchain to improve the federated learning framework can improve the reliability and security of the system and enhance data privacy. At the same time, with the expansion of network scale, how to optimize the network and allocate limited resources to provide efficient and high-quality services is a key problem that needs to be solved. BeFL can be combined with the economic incentive mechanism to comprehensively consider the number of digital twins and marginal servers, the batch size of training data and bandwidth allocation, and use the consensus mechanism to select the participating nodes in the joint training to improve the efficiency of the system. Existing BeFL studies have proposed reinforcement learning and asynchronous aggregation to improve operational efficiency, but further research is still needed [118−119].

4. Questions for further research

Although the integration of blockchain and federated learning can effectively alleviate some of the challenges faced by the latter, it does not completely solve the inherent efficiency, heterogeneity, game and security problems of federated learning systems. By summarizing the commonly used models, blockchain platforms, datasets, and evaluation criteria of BeFL architecture (as shown in Table 3), this section will summarize these problems for further research (as shown in Table 4) and give solutions in the existing literature.

Table 3 Experimental design of BeFL architecture

Table 4 Key issues and future directions

4.1 Efficiency of BeFL architecture

Compared with traditional federated learning, BeFL needs to consider the performance loss and time delay caused by the blockchain consensus process, which will lead to serious efficiency problems. Firstly, when the training node is a device that requires highly dynamic real-time performance (such as a smart car or a drone), its efficient consensus will inevitably face greater challenges in computing and communication resources. Secondly, although some deep learning models can be fused into the BeFL architecture, the deeper deep learning models need to store more learnable parameters for model fusion. Therefore, compared with the centralized federated learning method, the computational overhead of BeFL will still be limited. Finally, machine learning does not require strong consensus or consensus to converge, so the commonly used strong consensus protocols such as Practical Byzantine fault-tolerance (PBFT) are too strict for machine learning workloads. The commonly used models in existing studies are NN/CNN/MLP, but the number of model layers is shallow, the deployment of large models is difficult, and the training convergence speed is slow. Most of the blockchain platforms used are Hyperledger Fabric and Ethereum; The efficiency evaluation criteria usually include model accuracy, block generation time and running time. This section will summarize the three main methods of efficiency improvement in existing research, the main purpose of which is to improve the accuracy of model training, reduce communication and computational overhead, and optimize system operation and model training efficiency.

4.1.1 Adjust the block generation rate

In practice, the generation rate of data blocks will be affected by many factors, and it is necessary to comprehensively consider factors such as communication, computing power, and consensus delay to build and optimize the end-to-end model[92]. In terms of optimization methods, the end-to-end delay can be quantified by combining system dynamics, and the block generation rate can be dynamically adjusted under actual network conditions by considering the deviation between the system delay and the optimization target value in the BeFL architecture[120−121]. In addition, the combination of artificial intelligence technologies such as deep reinforcement learning to determine the block generation rate can also reduce energy consumption and improve system operation efficiency[68].

4.1.2 Compressing gradients and models

Gradient compression is a commonly used method to reduce the communication load in distributed machine learning. In addition to improving operational efficiency, gradient compression also protects privacy. In the BeFL architecture, a gradient compression scheme can be used to generate sparse but important gradients, and the gradients uploaded by edge nodes can be compressed to reduce the time required for communication, which can reduce the communication overhead without affecting the accuracy, thereby speeding up the training process of federated learning and further strengthening the privacy protection of training data[122−123]. Model compression is another research point of deep learning, which aims to reduce the complexity of the model and make the algorithm run more stably and efficiently. In order to improve the modeling efficiency in the BeFL architecture, model compression techniques such as knowledge distillation and federated learning aggregation algorithms can be used to implement model compression before parameter broadcasting in the blockchain network, so as to reduce the load of the entire blockchain network and improve the efficiency of system modeling[124].

4.1.3 Adopt a dual-chain architecture

On the one hand, the use of smart contracts to implement model aggregation and reward and punishment mechanisms on public chains will greatly slow down the federated learning process, and sending models on public chains such as Ethereum will lead to expensive costs for both the client and the server. On the other hand, the use of consortium blockchains (or private chains) is highly efficient for communication and allows only a group of authorized participants to view sensitive data, thus helping to solve data privacy concerns. Therefore, the hybrid architecture of public chain + alliance chain is adopted, and the aggregation is performed on the consortium chain and the reward and punishment mechanism is implemented on the public chain, which not only maintains the communication efficiency of the algorithm, but also reduces the operating cost[69, 125−126].

4.2 Heterogeneity of BeFL architecture

Unlike distributed machine learning running in the cloud, BeFL faces severe heterogeneity challenges. Since the participating nodes are allowed to join and exit the training process freely in the blockchain, and the types of participants are diverse, such as sensors, fog nodes, cloud service providers, etc., and the data feature dimensions when participating in the training are also different, this heterogeneity leads to the very diversity of update models provided and received by the training nodes, and the convergence performance of the algorithm will be poor. The dataset commonly used in existing studies is MNIST/CIFAR-10, while the complex heterogeneous dataset that can be divided into non-independent and identical distributions such as Federated-MINST\Shakespeare is rarely used. In the real scenario, heterogeneous problems at the data, model and network resource levels will be caused by different local data types and the complexity of participants.

4.2.1 Heterogeneous data

The data of the training model is usually non-independent and identically distributed, which is both a characteristic and a challenge of federated learning. There are many studies on data heterogeneity in federated learning, such as improving the training of non-independent homogeneous data by sharing a small part of data globally among all edge devices, and designing some optimization algorithms to ensure the stable convergence of the model when training with heterogeneous data. In order to solve the problem of data heterogeneity in the fault detection of industrial Internet of Things (IIoT), Zhang et al. [127] designed a platform architecture based on blockchain and federated learning, and proposed a centroid distance-weighted joint average algorithm. In order to generate an unbiased global model, the distance between the positive and negative classes of each client dataset is considered in the weight calculation, which solves the data heterogeneity problem of BeFL to a certain extent.

4.2.2 Model heterogeneity

Both federated learning and BeFL models need to train the main model or global model according to the data sources provided by all parties, and the local model is required to be isomorphic. However, in practice, each party trains the model based on its own local data, so heterogeneous models and federated learning based on heterogeneous models may be more common in practical applications. To this end, Wang et al. [128] proposed a blockchain-empowered decentralized secure multiparty learning (BEMA) system, which extends distributed multiparty learning to a decentralized structure, allowing each participating node to hold a heterogeneous local model and train it, which enhances the effectiveness and reliability of the BeFL architecture.

4.2.3 Heterogeneous network resources

Due to the time variation of the network environment and the heterogeneity of network resources, it is difficult to achieve stable, reliable and real-time interaction between edge devices and edge servers, which is especially obvious in the 5G ultra-dense network environment. To this end, Yu et al. [129] proposed an intelligent ultra-dense edge computing framework, which integrates blockchain and federated learning technology into the 5G ultra-dense edge computing network, and formulates a heterogeneous network resource and hybrid computing distribution mode for ultra-dense edge computing. In order to minimize the task execution time and the use of network resources, the framework optimizes the application partitioning, resource allocation and service cache configuration mechanisms in two different time frames, and the experimental results confirm the effectiveness of the proposed framework.

4.3 Game Problems of BeFL Architecture

Game theory is an important tool for designing BeFL architecture and mechanism. Firstly, game theory can be used to improve the consensus mechanism of BeFL. BeFL can design a consensus algorithm based on the theory of game and mechanism design according to the data quality of the nodes, the contribution and enthusiasm of the nodes to participate in joint training, etc. For example, the contribution of each node in joint training is measured according to cooperative game theory [46−47]. Incentive-compatible game analysis based on auction model to promote rational nodes to follow the protocol and maximize their profits[50]; According to the mechanism design theory, the actual task solving process of federated learning is integrated into the consensus algorithm of the blockchain (such as PoW)[45].

Secondly, game theory can make the BeFL architecture reach a certain equilibrium state under various constraints. For example, when a defense mechanism needs to be deployed to check if an adversary is attacking the BeFL architecture, the server will incur additional computational costs in addition to joint training. Generally speaking, different defense mechanisms have different effectiveness and cost for various attacks, which requires the combination of game theory to optimize the deployment of defense mechanisms. Game theory can also be used to simulate the interaction between devices and servers to determine the optimal conditions between computational overhead and communication costs, and to improve the block generation rate accordingly[95]. In addition, there is a similar situation between privacy protection and efficiency, for example, when differential privacy is adopted, higher privacy requirements need to add more noise to the data, and at the same time, the communication cost also increases, and game theory can be used to make a trade-off between privacy and efficiency.

Finally, because the BeFL architecture lowers the entry threshold for joint training, more users can be attracted to join the joint learning process, and game theory can be used to analyze the coopetition of user nodes in joint training[130−133].

4.4 Security issues of BeFL architecture

Although the introduction of blockchain technology has alleviated the security threats faced by federated learning to a certain extent, there are still many security challenges in BeFL itself. The existing literature mainly discusses the security issues of BeFL architecture, including node trustworthiness, data security and system security.

Node trustworthiness is very important for BeFL security construction. For example, adversary nodes can use poisoning attacks, inference attacks, etc., to destroy the global model and infringe on the privacy of other nodes, while honest nodes can contribute their own high-quality data and maintain the healthy and stable operation of the system. There are two main methods to ensure the trustworthiness of nodes in the research of BeFL architecture: one is to use identity authentication to ensure the authenticity of the identity of participants, and the common identity authentication methods mainly include the Kerberos protocol based on symmetric cryptography and the public key infrastructure (PKI) system based on public key cryptography, in order to ensure the uniqueness of the user's key identity under the mechanism of multiple verification. In order to alleviate adversary attacks and ensure the credibility of the data shared by nodes [134]; The second is to analyze the threshold of the proportion of adversary nodes, and to design the corresponding consensus incentive mechanism algorithm to ensure the security of the whole system[135].

In addition to the need for homomorphic encryption and differential privacy in the process of federated computing, data security is also reflected in the data security after the participants upload the data to the smart contract, the local model data security of the participants, and the storage security of the data on the computing device. In the study of data security of smart contracts, Kosba et al. [136] proposed a privacy-preserving smart contract development framework called Hawk. In Hawk, smart contracts are divided into private contracts and public contracts, and private data and related financial information are only visible to the contract owner after being written to the private contract. In addition, Zhang et al. [137] proposed a trusted data input system for Town Crier, in which the contract encrypts the request with the public key of Town Crier before sending the request, and Town Crier uses the private key to decrypt the request after receiving the request, so as to ensure that other users in the blockchain cannot view the request content. In this regard, FedIPR proposed by Li et al. [138] proposes a feature-based and backdoor-based watermark embedding and verification scheme, which aims to embed and verify private watermarks without leaking the private training data or watermark information of multiple parties. FedIPR is a technical solution to protect the ownership of federated models in a secure federated learning environment, which proves the validity, reliability and robustness of watermarks through theoretical analysis and experimental results, and helps to detect and eliminate free-riders in federated learning. Data storage security is usually achieved in the form of encrypted storage, using attribute-based encryption (ABE) and proxy re-encryption (PRE) to store encryption keys and control access to critical data, so as to ensure data storage security.

System security is mainly reflected in the security of computing environment, smart contract operation and communication. Computing environment security includes the isolation of the execution environment, the integrity of the application, and the confidentiality of data, etc., and the underlying hardware trusted execution environment such as Intel's SGX (Software guard extentions) is usually used to maintain the security of the computing environment[139]. The operation safety of smart contracts lies in the fact that the smart contracts that have been deployed on the chain are irreversible, and their potential security problems are difficult to repair once they are triggered. To this end, Chen et al. [140] proposed a smart contract high-burn operation detection tool called Gasper, which can automatically detect dead code, useless descriptions, and expensive loop operations. The communication security of the BeFL architecture lies in ensuring the confidentiality of communication and the integrity of the transmitted data in transmission, regardless of the use of homomorphic encryption or inadvertent transmission, and the common methods are to use onion router to protect the communication relationship between the two parties and user identity information such as IP addresses, and use the TLS (Transport layer security) protocol and DTLS (Datagram transport layer). security) protocol to ensure the security of network transmission and the security of uploaded data.

5. Future research directions

The deep integration of blockchain and federated learning will inevitably give birth to or empower some new models and new business formats, and derive new research opportunities and directions. In this section, we will briefly discuss the potential directions of BeFL-based data element market, game analysis and mechanism design, privacy and regulatory trade-offs, quantum computing, and the potential impact of the new generation of metaverse digital space on BeFL.

5.1 BeFL-based data element market

Whether for enterprises or governments, data is a very valuable digital asset, especially in the Internet industry, where the training of machine learning models relies on a large amount of data. However, with the advancement of data privacy protection regulations and the increasing importance of data privacy protection, data will remain on institutional and personal local devices instead of being shared in the cloud. In this case, the data element market will be an effective solution to promote the sharing and value circulation of data elements.

BeFL is expected to solve some of the technical challenges faced by the data element market. First of all, data is different from other assets, and when the data owner trades the data, the data owner loses ownership of their asset and cannot extract value from their asset in a sustainable way. Second, centralized data marketplaces rely on a trusted central entity to maintain data sharing, which can create data monopolies and violate user privacy. The data element marketplace based on the BeFL architecture can retain the privacy and ownership of data assets in the decentralized data market, and meet the information transparency and trust needs of buyers and sellers for the data assets to be traded[141].

At this stage, the main problems faced by the data element market are: which elements (data, models, computing power, etc.) will participate in market transactions; Under what circumstances will the matching of supply and demand between buyers and sellers occur; how to effectively price data elements in the transaction process; What kind of mechanism is used to ensure the fair and orderly operation of the market, etc. The existing literature has done preliminary research on these aspects.

Fan et al. [126] combined the advantages of public blockchain and consortium blockchain to propose a resource transaction system based on hybrid blockchain, which can reduce system latency in infrastructure. At the same time, the system adopts the reverse auction mechanism, and uses the payment channel technology to realize the trusted, fast, low-cost and high-frequency payment transactions between the requester and the edge node. Somy et al. [142] considered three types of market participants: data owners, model developers, and cloud owners in the market built by BeFL, and provided verifiable data in the blockchain system to resolve disputes. For the supply and demand relationship of the data element market, Ouyang et al. [143] considered that transactions may occur in the situation where multiple computing power owners need to access the same set of data, based on Ethereum smart contracts to coordinate multiple computing power owners, and automatically implement model scheduling and reward and punishment incentives.

5.2 Integration with next-generation AI models

The integration of large-scale intelligent models and algorithms based on the BeFL architecture will become an important development trend for the application of a new generation of artificial intelligence, and lay a solid foundation for data and trust for the rapid development of the Web3 ecosystem in recent years[144]. At present, the main combination points include the intelligent recommendation model based on BeFL, the new generation search model based on BeFL, and the artificial intelligence generated content (AIGC) model based on BeFL.

The intelligent recommendation model based on BeFL aims to solve the problem of weak privacy protection caused by the centralized architecture of the current recommendation system, and uses the BeFL architecture to authenticate the identity of participating users and select recommendation clients to ensure the credibility of recommendation items[145−146]. According to the BeFL privacy protection mechanism, the client collects local user behavior data (such as web click data and favorite data) and trains the model collaboratively under the restriction that the data does not leave the local area, so as to build a joint recommendation model to solve the technical problem of poor recommendation effect under the condition of protecting data privacy.

The search model based on the BeFL architecture also has a wide range of requirements: on the one hand, it is necessary to build a decentralized and strong security and privacy joint search model for the scenario where user data is stored in a local server. On the other hand, with the wide application of blockchain technology, there are also business needs such as data query and statistical analysis between different users and blockchains. Therefore, based on the integrated search model of BeFL architecture, the encrypted text of the stored files and the search index of the blockchain foundation can be used to filter and verify the participating nodes, the smart contract can be used to search the encrypted data to achieve secure and trustworthy search, and the federated learning technology can be used to process heterogeneous data and analyze the search results to meet the personalized needs of users for search services[52, 147].

AIGC is a new type of content creation after professional content production and user-generated content, and its development is inseparable from users' high-quality data contribution and active participation. In the AIGC model based on the BeFL architecture, blockchain technology can be used to return the ownership and control of network data to users, generate digital identities and data assets highly attached to users, and design economic incentive mechanisms to attract users to participate and contribute data. At the same time, the use of federated learning technology to process user data and adapt to AIGC model training to improve the model effect will be one of the trends in the development of AIGC models[148−149].

5.3 Privacy and Regulatory Trade-offs

Data privacy protection is essential for businesses and individuals. The GDPR promulgated by the European Union adheres to the premise of strong data protection, but there are regulatory blind spots in data use and the security challenges that come with it. Therefore, it is necessary to find a more appropriate balance between data privacy protection and business process supervision, so as to achieve maximum compliance utilization of data while ensuring data security.

In practice, Internet companies and big data companies are stronger than government agencies in terms of data computing capabilities, so data regulation and governance usually require a multi-party participation and collaborative and trusting environment. How to realize the privacy protection of all parties' data in such a multi-party collaborative supervision environment, and on this basis, realize the active sharing and incentive of regulatory data, the trusted integration and conflict negotiation of cross-departmental regulatory business, and the cross-chain supervision coordination and supervision services for the multi-party weak trust environment, so as to achieve an efficient balance between multi-party active collaborative supervision and privacy protection, are the key technical problems that need to be solved urgently.

Therefore, it is necessary to combine blockchain and privacy computing technology represented by federated learning to meet the needs of collaborative supervision and data sharing in multi-regulatory subjects and weak trust environments[150], design cross-industry, cross-departmental and multi-party regulatory norms and conflict negotiation techniques, study the design methods of active sharing and incentive mechanism of regulatory data, and design the privacy protection mechanism of regulatory data, so as to support active and credible regulatory collaboration and data sharing.

5.4 Opportunities and challenges brought by quantum computing

Quantum computing is a new computing mode that follows the laws of quantum mechanics to regulate quantum information units for calculation. The existing quantum computers are better than the traditional computers in terms of computational efficiency, and the quantum training method is used to replace the classical machine learning algorithm in terms of computational efficiency and security. At the same time, quantum data privacy and transmission efficiency are issues that should be considered when conducting large-scale quantum machine learning. Combined with federated learning, a new computing paradigm, it can effectively solve the problems of quantum data privacy and transmission efficiency[151−153]. But inevitably, quantum federated learning also faces traditional problems such as central node failure.

Therefore, with the development of quantum computing, combining blockchain and federated learning technology, a blockchain-based quantum federated learning computing framework is designed, which can not only enable quantum machine learning models to perform distributed quantum learning among quantum clients without transmitting quantum data itself, but also effectively resist central node failures, improve the collaboration efficiency of distributed quantum computing resources, and accelerate the application process of quantum computers.

At the same time, quantum computing is a potential threat to the underlying cryptographic system of the blockchain. With the acceleration of the development process of general-purpose fault-tolerant quantum computers, the traditional hashing algorithm and asymmetric encryption algorithm based on mathematically difficult problems are facing the risk of halving the attack difficulty or completely defeating it. Therefore, researchers have proposed quantum-resistant blockchain models based on quantum-resistant cryptographic algorithms such as coding problems and lattice problems, as well as quantum blockchain models based on quantum cryptography such as quantum key distribution and one-time-one-secret to deal with the potential threats of quantum computing. Generally speaking, the quantum-resistant blockchain model is more suitable and has been gradually integrated into the existing public chain system, while the quantum blockchain model is more suitable for the consortium chain system with quantum capabilities and fixed nodes.

5.5 Future Digital Space - Applications in the Metaverse

With the rise of the concept of the metaverse, blockchain and federated learning, which are its important technical foundations, have received widespread attention. The decentralized trust mechanism of blockchain and the privacy protection characteristics of federated learning can effectively solve the problems of identity fragmentation, data isolation, and loss of rights and interests in the digital world, and are the cornerstones of building an open, fair, and credible metaverse. Therefore, the BeFL architecture is helpful to realize data collaboration and value exchange between different metaverse platforms, applications, and spaces, and solve several key problems in the metaverse vision[154].

First of all, the BeFL architecture can empower users to manage their digital identities and assets autonomously. In the metaverse, users need a unified digital identity that can be cross-platform, application, and space, as well as digital assets that can be created, confirmed, and circulated, such as virtual land, game equipment, artworks, etc. The identity and management mechanism in the traditional Internet has the problem of centralization and fragmentation, and it is difficult for users to truly own and control their digital identity. The BeFL architecture can use blockchain to provide users with a decentralized ledger system, allowing users to establish and verify their own digital identities using privacy-preserving computing technology, and record and trade their digital assets in the form of non-fungible tokens (NFTs) [155].

Secondly, the BeFL architecture can realize the privacy protection and trust guarantee when user data collaboration and interaction in the metaverse. Users need to provide a lot of personal data in the metaverse, such as location, preferences, behavior, etc., to achieve a more realistic and immersive experience. Sharing this sensitive data directly exposes user privacy. The BeFL architecture can provide users with a privacy-preserving data collaboration and interaction system, allowing users to train machine learning models on local nodes and exchange only model parameters, rather than raw data. In addition, the BeFL architecture can provide users with a decentralized trust mechanism, which supports users to use cryptography mechanisms to verify and record the exchange process of model parameters, preventing tampering and fraud[156−157].

Finally, the BeFL architecture can build a new economic system and community governance structure in the metaverse. Users in the metaverse need an economic system that can measure their contributions, as well as a community governance structure that can participate in and influence the direction of the metaverse. Blockchain can provide users with a value exchange system, support users to use cryptocurrencies for payments and rewards, and define and enforce various rules and protocols in the form of smart contracts. Federated learning can provide users with a data collaboration system, support users to use machine learning models for knowledge sharing and innovation, and achieve democratic governance in the form of DAOs [158−161].

6. Conclusion

The integrated innovation of blockchain and federated learning is one of the important trends in the development of a new generation of information technology. In this paper, the conceptual model of BeFL is proposed, the basic workflow of BeFL is described, and the key research problems and existing research progress in this field are discussed from six dimensions: infrastructure, consensus mechanism, economic incentives, smart contracts, privacy protection and application fields. In this paper, we also discuss the open research problems and application scenarios of BeFL. It should be pointed out that the combination of blockchain and federated learning is still in its infancy, and it is facing new development opportunities and severe research challenges, which need to be discussed in combination with more research fields and application scenarios. It is expected that this paper can provide useful reference for future research.

Disclaimer: The articles and pictures reproduced by the official account are for non-commercial educational and scientific research purposes for everyone's reference and discussion, which does not mean that they support their views or confirm the authenticity of their content. The copyright belongs to the original author, if the reprinted manuscript involves copyright and other issues, please contact us immediately to delete.

Note: If you can't display the full article, you can V search for "Artificial Intelligence Technology and Consulting" to view the full article

Artificial Intelligence, Big Data, Multimodal Large Models, Computer Vision, Natural Language Processing, Digital Twins, Deep Reinforcement Learning······ You can also add V "Artificial Intelligence Technology and Consulting" to sign up for the course