In the past year, from the CCTV 315 evening party exposing the illegal use of face recognition cameras in many offline stores to analyze consumption behavior, to some community residents encountering the helplessness of "not brushing their faces and not going home", to the continuous reduction of the threshold for the use of technologies such as "deep forgery", face recognition applications have faced many challenges.
On December 17, the "2021 Woodpecker Data Governance Forum" hosted by the Nandu Personal Information Protection Research Center was held in Beijing. The Nandu Artificial Intelligence Ethics Research Group released the "Face Recognition Application Scenario Compliance Report (2021)" (hereinafter referred to as the "Report") at the meeting, which evaluated and analyzed the compliance of 20 mobile face recognition applications.
The report shows that 60% of apps with face recognition functions do not have separate face recognition rules. From the perspective of specific scenarios, the overall performance of the payment transfer class is better, while the app in the fields of entertainment special effects and access control is relatively lacking in rule notification. In addition, after technical evaluation, the 4 entertainment special effects Apps tested did not encrypt the transmission of personal information, and the face picture link could be publicly accessed, which posed a greater security risk.
Sixty percent do not have a separate face recognition agreement, and the "notification and consent" of brush face payment is more compliant
On July 3 this year, a "face recognition must wear clothes" entry was hotly searched. Many users mistakenly believe that the face recognition system will only upload information about the face part, so they may be face recognition while lying in the quilt, taking a bath, and hugging the other half. As everyone knows, the parts other than the face are also recorded by the camera and seen by others.
Behind this embarrassment reflects that while the Face Recognition function of the App is popular, there is a certain difference between the way it collects personal information and the public perception. Combined with the ranking of App downloads in multiple Android app stores and user complaints on the Internet open platform, the research group selected 20 apps such as payment transfer, real-name verification, property access control, and entertainment special effects for evaluation.
The results show that only 8 apps separately prompt relevant rules before users use the face recognition function, including "Cloud Flash Pay", "Industrial and Commercial Bank of China", "Alipay", "JD Finance", "Taobao", "JD", "Agricultural Bank of China" and "QQ".
Specific to the scene category - in the face brush payment scenario, the tested App will display face recognition rules to the user when he opens the face brush payment. For example, the General Rules for Biometric Services of "Alipay" and "Taobao", the "Jingdong Face Service Agreement" of "JD" and "JD Finance", the "Face Recognition Service Agreement" and "Face Payment Agreement" of "Cloud Flash Pay", and the "Brush Face Payment Business Agreement" of "Industrial and Commercial Bank of China".
In the one-time real-name verification scenario, only the three apps "QQ", "Agricultural Bank of China" and "Taobao" show the face recognition rules.
None of the 4 apps in the entertainment special effects space has separate face recognition rules. However, "Fun Show" and "ZAO" will have a short prompt for the processing rules of face information, for example, "Fun Show" will pop up a prompt when the user triggers the "AI Dressup" function, telling "Face photos will be deleted immediately after the video is synthesized, and your face photos and data will not be retained".
The pop-up window of "Fun Acting" informs
In the two property access control apps, "Hitomi Community" did not show any agreement before entering the face; "Relatives Open doors" will pop up the "Relatives and Neighbors Face Open Door Service Agreement". However, the agreement is not specifically for the processing of face information, and does not specify the rules for the processing of face information.
The report shows that in the 8 apps that provide separate face recognition rules, there are also situations where the prompts are not in place. For example, when the three Apps of "Alipay", "Taobao" and "Industrial and Commercial Bank of China" opened the face payment function, they did not obtain the explicit consent of the user.
In addition, the "Industrial and Commercial Bank of China" has a variety of face brushing scenarios such as payment transfer and real-name verification, but some scenes prompt face recognition rules, and some scenes do not prompt. Specifically, when the "offline merchant face payment" function is set up, the App will display the "Face Brush Payment Business Agreement" to the user; when using the "Cloud Custody" (a cloud storage function), although the user is required to log in with the face, this link will not show the user any face recognition rules - the user can click a button to directly enter the face verification.
2) The gap between face recognition rules is large, and the storage location and time limit are not clear
The report shows that unlike the privacy policy, the framework of the face recognition protocol of different apps is different, and its level of detail and content are also quite different.
For example, the QQ Face Recognition Function Service Agreement has only three short paragraphs, and does not explain the storage period, storage method, processing rules and other information of face information. The "Cloud FlashPay App Face Recognition Service Agreement" informs the specific steps to close the face login service and the solution when face recognition fails, and indicates that the user's information will be protected in accordance with the privacy policy, and the user will be informed and consent will be sought when the agreement is updated.
The evaluation results show that many App face recognition rules do not inform the storage time limit or location, and only 6 Apps mention the storage of face information.
The four apps of "JD.", "Taobao", "Agricultural Bank of China" and "JD Finance" said that they would save face information within the "required time limit"; "Taobao" further promised that the original image of the face would be deleted in time after completing the verification service. In contrast to Taobao, the Industrial and Commercial Bank of China said that images taken during each verification may be saved to help correct the algorithm.
Regarding the storage location, only the biometric information entered by the "Alipay" 1 app promises to be stored locally on the device - "The biometric information you enter will only be saved on the device, and once you change the device, you need to re-enter the biometric information on the new device".
3) The entertainment special effects app face picture link can be publicly accessed
The report also used technical means to conduct data security tests on 20 apps. Data security detection detects the collection and network transmission of personal information in the real environment of the relevant application by installing and activating the tested App, logging into the physical account and triggering the face collection and upload function, and using reverse analysis, data capture and other technical means.
The evaluation results show that 16 of the 20 apps encrypt and transmit encryption of personal information, and 4 entertainment special effects apps have problems.
For example, "Fun Show" does not encrypt the face information. The App's "AI Dressup" function is to generate a face-changing video by uploading a photo and then selecting a video template. However, due to the lack of encryption measures, the link to the user's face-changing video can be publicly accessed. This means that the face-changing video may be accessible to anyone.
The face-swapping video of "Fun Acting" is publicly accessible
Three apps, such as "ZAO", "More Beautiful", and "New Oxygen Medical Beauty", although they use the HTTPS secure transmission protocol, do not encrypt the data itself, resulting in the user's face photos and other information being uploaded to the server, and the server returns a link that can be publicly accessed by the Internet. The research group copies the relevant links to the browser, and can directly view the corresponding information. This means that once an attacker intercepts a transmitted packet, it will gain access to a range of sensitive personal information from the user.
The report believes that face recognition application software shows a phenomenon of unevenness in terms of rule notification and technical security, and the application of different scenarios has obvious gaps in compliance. Head banks and Internet platform companies are more standardized in rule making, but there are still obvious problems in policy transparency, while some entertainment special effects apps have very obvious security vulnerabilities, which may become the "hardest hit area" of privacy leakage in the field of face recognition, which should attract the attention of developers.
Wen/Nandu Artificial Intelligence Ethics Research Group researcher Li Yaning Hu Gengshuo