Shulou(Shulou.com)06/03 Report--

This paper provides ten face recognition policy recommendations to deal with the potential face recognition risks from three aspects: face data, testing technology, authentication and evaluation subjects.

Core summary of the top ten recommendations:

Limit data storage period, limit data sharing, set up face recognition signs in public places to improve face recognition accuracy, carry out third-party independent evaluation, reduce incidental information collection business scenarios into opt-in and opt-out mechanisms to develop unified technical standards, improve standardization organization certification to ensure data representation and testing practice

At present, the application of face recognition continues to go deep into daily life, and it is widely used in the fields of missing search and rescue, security upgrade, blind guide, anti-terrorism and so on, but it also causes public unease and doubt.

In view of the fact that there are still many uncertain risks in the application of face recognition technology, this paper provides ten face recognition policy recommendations to deal with the potential face recognition risks from three aspects: face data, testing technology, authentication and evaluation subjects.

1. Limit the duration of data storage

The so-called face recognition (Facial Recognition) refers to the static image or dynamic video of a specific scene, using a number of existing stored face image databases to verify and recognize the identity matching of single or multiple people in a specific scene.

Face recognition usually consists of three parts: face detection in specific scenes (Face Detection) and face segmentation, extraction and analysis of facial features, and matching database to recognize other people's faces.

The long-term digital storage of face images is one of the most frightening reasons for face recognition. There are many risks of abuse of these face information, and there is an urgent need to change the way to resolve public anxiety. One of the ways is to set the storage period of images and videos. When the emergency situation is in a sudden state, it is necessary to store some specific image data.

Once the crisis period has passed, there is no need to retain the face data retained in the critical moment. Therefore, for most application scenarios, limiting the storage period of face data can balance the multiple benefits of face recognition and minimize the risk.

The specific storage time varies from situation to situation. For example, specific images compiled in response to emergencies have high instantaneous value, while other cases need to be included in a huge database for subsequent matching and recognition of other face features.

This paper introduces a machine learning model-federated learning (Federated Learning), which ensures the security of data. This learning mode is a decentralized machine learning scheme for training data, which ensures that the data is only stored in the camera terminal and not transmitted to the central data center, thus improving data security.

two。 Restrict data sharing

It is worrying that the same data is used to share and transfer among many different purposes. For example, the US vehicle Administration sells recognition images to third-party agencies for face recognition in other scenes.

This worry is that in the process of sharing face data, the data subject is completely unaware that the data is being used for other purposes, resulting in the failure of the informed consent mechanism, even if the data subject knows in advance. It is also likely to disagree to share facial recognition data with commercial organizations for commercial use. Therefore, if you want to share face recognition data across scenes, you must provide legitimate reasons.

According to the Brookings Institution poll, people's attitudes towards face recognition vary from scene to scene. Among them, the consent rate of using identification technology to protect school students is as high as 41%, the consent rate for airport security checks and stadiums is about 30%, and the lowest consent rate is used by stores to prevent theft.

3. Set up face recognition signs in public places

Whether it is a private body or a public authority, when it is used to take photos, shoot videos or collect public information for other purposes to recognize other people's faces, they should set up clear face recognition logos in public places to clearly inform the public that there is a face recognition system here, thus potentially affecting the public's compliance with public order.

In the long run, it will not only enhance public safety awareness of the use of face recognition, but also protect the freedom of choice of groups who do not want to be recorded facial information.

4. Improve the accuracy of face recognition

The accuracy of face recognition will first be affected by the recognition object-different ethnic groups. Facial pattern features can be divided into two categories: skin color features and grayscale features. as important information of human face, skin color is independent of other facial details and has relative stability. However, the recognition accuracy of face recognition system for whites is higher than that of non-white groups, and the darker the skin is, the lower the accuracy is. At this time, the tendency of recognition bias will be magnified by the relative stability of skin color.

In addition, due to the incomplete and non-representative training data of ethnic groups, it will also aggravate the inherent bias of face recognition. When the technology is applied to law enforcement, border security, retail, airports and other public places, there will be prejudice and discrimination against different ethnic groups.

Another important factor contributing to the decline in recognition accuracy is lighting. After all, as a three-dimensional object, human face is inevitably disturbed by external factors such as light shadow, irradiation intensity and so on. Illumination will change the relative distribution of face image grayscale, so the change of face image caused by illumination is higher than that caused by individual differences.

According to a study by Cardiff University (Cardiff University), there have been thousands of face matching errors in Australia. From this, it can be seen that it is most urgent to define the accuracy standard of face recognition before it is applied in public places on a large scale, and the clarity of the standard can be judged by the degree of influence on people's lives. If the law enforcement process seriously interferes with the core rights of the people, such as being arrested or imprisoned, the degree of recognition must reach the corresponding level.

5. Carry out third-party independent evaluation

The introduction of third-party independent evaluation can boost public confidence in their face recognition products and services, and consumers want to buy products to play their inherent role without causing other technical problems. In order to help consumers understand the functions and potential dangers of face products, we can consider establishing a star rating system for face recognition, or imitating the Energy Star Program (Energy Star) jointly implemented by the U.S. Department of Energy and the Environmental Protection Agency to include many face recognition applications in the scope of third-party evaluation certification.

6. Reduce incidental information collection

Some face recognition applications will collect huge amounts of information that have nothing to do with the main purpose, which violates the "minimum adequacy principle". For example, when the police take a personal camera to the scene to inspect, they are not only filming the suspect, but also happen to photograph nearby passers-by. Unless there is a clear indication that the evidence is relevant to the case, law enforcement agencies do not need to retain irrelevant information. When the captured image is no longer of investigative value, it can be blurred or even deleted.

7. Business scenarios are integrated into opt-in and opt-out mechanisms

The so-called opt-in refers to the need to obtain the consent of the data subject before the facial biological information related to recognition is shared. For example, when the face recognition technology associates the detected name with the personal portrait and pushes the commercial advertisement to the person, it is necessary to respect the right of consent in advance.

Under the background of strengthening personal data protection around the world, data subjects are more and more concerned about personal privacy. As the privacy of biological information behind face recognition, it has typical identifiability and belongs to personal biological data, which should be included in the category of personal sensitive data and should be protected.

In addition to the opt-in mechanism, the opt-out mechanism (opt-out) and the right to be forgotten can also be applied. In situations where the risk coefficient is low and there is no need for long-term data storage, people are given the right to choose that the relevant institutions should not continue to collect or maintain data sharing at an acceptable and reasonable level. With the passage of time, the high-value face data may gradually become outdated, irrelevant, out-of-range, and even harmful, so it is very necessary to introduce the right to be forgotten, which can enhance the public acceptance of face recognition.

8. Formulate a unified technical standard

It is a common means for market private entities to formulate technical standards to ensure product safety. Take the mobile communication technology of that year as an example, when it is still in the stage of development, industry experts have formulated general standards for communication, security and compatibility. All mobile phones must meet the above technical specifications before they can be sold. Today's face recognition technology follows the same principle.

Face recognition technology should also establish international general technical standards to ensure the safety of face technology, privacy will not be violated, and ease the fear of the world. Just as all technologies have two sides, face recognition technology is also in urgent need of creating "responsible face recognition".

Fortunately, the American Electrical and Electronic Engineering Association (IEEE) and the National Institute of Standards and Technology (NIST) are developing unified technical standards to regulate the application of related technologies.

9. Improve the certification of standardization organizations

The safety factor of enterprise system is verified by International Organization for Standardization (ISO). Enterprise-specific products are assessed by ISO to see if they meet the requirements of regulatory rules, and compliance tests are carried out by third-party agencies, so as to protect consumers' right to know the technical standards of the whole process.

In the United States, NIST is responsible for product technology certification. It compares face detection results through a public database and authenticates related applications at the same time. However, there are voices criticizing that NIST relies too much on the initial data on private websites and can not be extended to daily use scenarios; NIST data selection is too narrow, only focusing on face data related to law enforcement; testing standards rely solely on image quality and operational functions, and so on. Therefore, face technology verification should combine automatic testing with manual verification, improve standardization organization certification, and create reliable testing and credible verification.

10. Ensure data representativeness and test practicality

In order to ensure the accuracy of face recognition, facial verification, technical standards and government compliance testing need to be based on widely representative, non-specific-purpose data. Under the tide of commercialization of face recognition, it is particularly important to use tabular database for baseline testing and product authentication.

Single-purpose data, such as police facial photos, do not fully represent all groups, and the value of the test will be greatly reduced. In addition to the need for representative data, the practice of testing is also particularly important to raise public doubts about face recognition. After all, the face recognition test based on massive image information, practical application environment and representative population sample grouping can effectively overcome the negative impact of lighting conditions and image resolution on test accuracy.

Note: this research report is from the 10 actions that will protect peoplefrom facial recognition software published by the Brookings Institution (Brookings Institution) on October 31st.

Translator: Cai Xiongshan, Yuan Jun; official account: Tencent Research Institute (ID:cyberlawrc)

Source: https://mp.weixin.qq.com/s/A8TII44wUTQi_cmu2ztvsA

Https://blog.csdn.net/weixin_42137700/article/details/103553310

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.