background
Based on the world's leading face matching algorithm, Ant Financial has developed interactive face live collection and recognition technology and image desensitization technology, and achieved a highly concurrent and reliable system security architecture through Ant Financial Cloud. Based on this, the core products of face verification have been successfully commercialized and widely used in authentication scenarios such as Alipay and online merchant banks. In the latest test report, the accuracy of face recognition by this technology has reached 99.6%, and the accuracy is 99.99%, which is far beyond the accuracy of naked eye recognition of 97%.
Since Alipay introduced face recognition technology, it has been widely used as the main authentication method in user login, real-name authentication, password retrieval, merchant audit, payment risk verification and other scenarios. Since it was launched at the end of 20 15, it has served more than1500,000 users. Compared with the traditional password-based authentication method, face recognition technology has greatly improved in security, reliability, recognition performance, user experience and so on, which is of practical significance to realize the internet financial scene.
InfoQ interviewed Jidong Chen, the head of biometric technology of Ant Financial, about the face recognition technology of Ant Financial.
Dr. Jidong Chen is currently the director and senior expert of Global Nuclear Platform Department of Ant Financial Services Group, responsible for the research and development and application of biometric technology. Lead the team to successfully launch face recognition technology and apply it on a large scale in financial scenarios such as online merchant bank and Alipay, so that Ant Financial can maintain the global leading position in biometric intelligent technology and application. Jidong Chen used to be the chief data scientist of Renren Game Big Data Research Center and the director of the Big Data Lab of EMC China Research Institute.
MIT Science and Technology Review released the list of global breakthrough technologies in 20 17, among which the face-brushing payment list is famous, and Ant Financial is the representative company of this technology. As the person in charge of biometric technology of Ant Financial, what do you think of this matter? Does this also mean that China's face recognition technology is in the leading position in the world?
Generally speaking, the list description is more accurate. The breakthrough technology on the list is face payment rather than face recognition. Breakthrough is defined as "bringing people high-quality scientific and technological solutions". So it is not only the technology itself, but also the application scenario and how to apply technology to change people's lives. Emphasizing the application of technology is also the orientation of the research and development of ant financial service technology. In addition to technology itself, we are more concerned about how the application of technology can bring people equal financial services. Identity recognition and authentication are the foundation of all financial services, so online identity authentication based on face recognition plays an important role in digital inclusive finance service.
Computer vision technology represented by face recognition has made great progress in the past few years, mainly due to the in-depth application of deep learning technology, the enhancement of computing power and the explosion of massive data. However, face recognition technology has only been fully commercialized in the past two years, and face authentication and face payment are also in their infancy, and there are many new scenarios that can be applied.
On 20 16 Yun Qi Conference, the Ant Financial Exhibition Area opened the "Future Cafe", where guests can complete the payment by brushing their faces in front of the camera. It is reported that the face payment is about to land in the real scene.
In the field of face recognition research, a group of outstanding China researchers are an important force to promote the continuous development of technology. It can be said that in the world, China's face recognition technology is in a leading position in both technology and application. Face Brush Payment is jointly developed by Ant Financial and China Face Recognition Technology Company. Ant Financial has developed a patented live detection technology based on the core comparison algorithm of face recognition, and combined with its multi-dimensional core technologies such as risk control and anti-attack security strategy based on financial cloud, it can provide financial-grade accuracy and security.
Can you talk about the investment and development of Ant Financial in the field of face recognition in recent years? What is its position in the whole field of face recognition?
Ant Financial began to invest money and talents in the field of face recognition several years ago, and has also been continuously investing in the research and development of other biometric technologies. Biometric technology has become an important part of ant financial service technology system and safety risk control system. Ant Financial began to apply face recognition technology to Alipay user login, real-name authentication, password retrieval, payment risk verification and other scenarios on 20 15. Up to now, it has been used by more than1500,000 users. As far as we know, it is the face recognition system with the largest number of users and visits at home and abroad, and it is also the first large-scale commercial online system in the financial field.
Face Recognition Technology of Ant Financial Service
Can you briefly introduce the commonly used algorithm models for face recognition? What kind of algorithm strategy has Ant Financial adopted?
Generally speaking, face recognition refers to face matching algorithm, which is divided into 1: 1 verification and 1:N recognition. The core of the algorithm is to let the machine automatically judge the similarity between different face images. The research of face recognition system began in the 1960s, and has been continuously improved with the development of computer technology and optical imaging technology since 1980s. The algorithm models involved in this technology development include recognition models based on local facial feature points, recognition models based on global feature transformation or geometric features, and recognition models based on 2D or 3D template modeling. At present, face recognition technology has turned to a recognition model based on Convolutional Neural Network (CNN).
In addition to the accuracy of the recognition model, there is also an important link in face recognition, which is to ensure that the face picture to be recognized by the machine comes from a living face, not a fake face such as a photo, video or mask. Therefore, in vivo detection technology is also the key to the successful application of face recognition. In vivo detection involves many algorithms, which are also closely related to sensor technology. For example, fingerprint recognition can detect the living body through capacitance/inductance sensor, and iris recognition can detect the living body through infrared camera. Because the popularity of infrared cameras on smart phones is still very low, the current living face detection technology mainly relies on a series of software algorithms, including recognition models based on action interaction and recognition models based on image analysis.
Ant Financial is advancing in face recognition and in vivo detection, and is developing multi-factor biometric technologies such as eye pattern recognition and voiceprint recognition to enhance face recognition. In addition, an intelligent identification model based on user behavior and different scenarios is developed based on big data analysis technology, thus forming a complete identity identification solution.
Can you introduce the peak amount of face recognition requests in Alipay's face-brushing payment service? What order of magnitude can the number of requests per day be? What kind of technical architecture does Ant Financial adopt to support its business?
The technical requirements and difficulties of financial face recognition are summarized as follows:
1. High security: face detection technology (to prevent photo forgery, video, masks, professional software tools and other attacks).
2. High accuracy: In the case of extremely low false recognition rate (
3. High availability: large-scale concurrent face matching service (tps = ""> 1000)
4. High real-time performance: face matching results are returned in real time (response time
Alipay's face recognition needs not only financial accuracy and security, but also high stability, reliability and low real-time response. Based on the infrastructure of Ant Financial Cloud, a highly available and dynamically expanded service framework system is realized to ensure that the face brushing service can meet the needs of high concurrent peaks such as Double Eleven and New Year Red Packet.
How does Ant Financial Service do data backflow? Can you introduce it to us?
Data reflow is indeed an effective means to improve the accuracy of algorithm recognition and improve the user experience of products. Under the premise of strictly observing the data security and privacy protection of Ant Financial, we verify and analyze the robustness of face recognition in various real environments by recording some key parameter information (such as illumination, distance, angle, duration, etc.). In the process of user's face brushing, the algorithm and product are further improved according to the analysis results in this real scene, so as to realize a completely data-driven closed-loop product development and optimization.
Difficulty of face recognition
Can you introduce the data level of Ant Financial Face Recognition Database, the time-consuming from face detection to returning comparison results, and the accuracy of face recognition? What is the applicable scope of accuracy? Is the accuracy only for Han people or all races (ethnic minorities, whites, blacks)? Is there any difficulty in face recognition of different races? How to solve it?
At present, Ant Financial is only aimed at Alipay real-name users of Chinese mainland citizens. Up to now, one-third of Alipay's 450 million real-name users have used the face-brushing service to log in, authenticate their real names, retrieve their passwords or authenticate themselves in high-risk transactions. The pass rate of face recognition (such as face-brushing login) is over 95% (a large proportion of failed users quit voluntarily). The faces of different races have greater diversity, which will challenge the accuracy of face recognition system. However, at present, the recognition model based on deep learning has the possibility to deal with massive data. If the face data of different races can be continuously trained and learned, this problem can also be solved well.
Can you talk about the development status and difficulties of face detection, in vivo detection, image desensitization and face comparison respectively? How to wear glasses, masks, masks, or brush your face with photos and videos?
Face detection: Face detection algorithm is the most mature branch of face recognition technology at present, and its accuracy and lightweight have met the needs of commercial use. Besides serving as a background service, it is also widely used in front-end devices such as smart phones and digital cameras. At present, the challenge is face detection in low light environment and large angle side face conditions.
In-vivo detection: In the past few years, in-vivo detection technology has also made great progress, which has been able to solve most photo and video attacks, but the means of in-vivo attacks are also evolving, especially the faces synthesized or transformed by various face-related modeling software are becoming more and more realistic. Face detection technology will be a process of continuous attack and defense and continuous improvement.
Image desensitization: image desensitization will bring information loss, which is contradictory to improving the accuracy of face recognition. Ant Financial has developed a unique one-way conversion desensitization technology, which can solve this problem well. At present, there are not many academic achievements in this field.
Face matching: At present, the recognition ability of the machine has surpassed the human eye, but low light, exaggerated expression, heavy makeup, aging and twins are still problems that need to be solved continuously. With the continuous accumulation and training of data, the performance is also improving.
challenge
Can you talk about the biggest challenge in the field of face recognition? Can you talk about their respective challenges from the perspective of face recognition algorithm and engineering?
In the algorithm, we also need to improve the accuracy of face recognition and in-vivo detection. As mentioned above, identifying problems requires guarding against evolving and emerging new attack methods.
The engineering challenges mainly lie in the user experience, the stability and reliability of the system, and constantly lowering the user's use threshold while ensuring the ultimate face-brushing experience, involving interactive copywriting, device compatibility, algorithm acceleration, parameter adaptation and many other aspects. Because the core of face recognition is image feature extraction and comparison, which is a CPU-intensive computing application, facing the authentication needs of hundreds of millions of Alipay users, especially in the case of high concurrent peaks such as Double Eleven and New Year red envelopes, how to ensure the performance and high availability of face brushing service is a systematic challenge.
Ant Financial Face Recognition product was officially launched in July 2065438+2005. Prior to this, all products were tested on a small scale, and rapid product optimization and iteration were carried out. We found the real scene very complicated. Users will brush their faces at different angles and postures indoors and outdoors in all kinds of light, day and night. Some will lie in bed and brush their faces, and some will even brush their faces when applying masks. How to solve the face brushing experience in various complex realistic environments, especially when the number of users reaches 100 million, is a great challenge. This is not only an algorithm problem, but also involves a series of problems from product, interaction, user experience, environmental parameter adaptation, security policy, high concurrency system architecture and so on. This is a systematic project. After more than a year of product optimization, we can now ensure a good face brushing experience and safety in various complex environments.
Future goals
Can you talk about where face recognition may be applied in the future besides face payment?
Certification has become the infrastructure of Internet finance, even in internet plus. The authentication method based on face recognition can better prove that "you are you" in the digital world and improve convenience and security. In addition, the credit system is the basic service of the whole society, and identity recognition and authentication are the basis of the credit system. The core foundation of all credit services is to know the personal credit rating, of course, the premise is to prove that "you are you", otherwise the evaluated credit rating may become someone else's. In addition to the application of credit and finance, the security field is also an important application of face recognition. In many cities across the country, train stations have put face recognition tickets, and Beijing airport has cleared customs.
Can you introduce the stage and goal that Ant Financial Services hopes to achieve in the field of face recognition in the next few years?
First of all, keep the world leading in technology, drive the deeper application of various business scenarios, and form a virtuous circle of AI-driven and data-driven. At the same time, not only in China, but also in the world with the internationalization of Ant Financial, providing more users with safe and convenient face brushing products and services.
How to treat the development of the whole face recognition industry in the next few years?
At present, the face recognition industry as a whole is still in its infancy. As mentioned above, there is still a long way to go before it can be applied to all walks of life and various user groups. There is no industry standard for face recognition both at home and abroad. At present, there is still a certain threshold for the use of products related to face recognition, which has not reached the level of popularization. However, with the continuous investment of technology, the continuous maturity of industrial environment and the continuous introduction of relevant standards, I believe that the next few years will usher in a real explosion of face recognition industry applications.
notification
What advice do you have for newcomers who want to switch to face recognition? Is the threshold too high to cut in?
Face recognition is a systematic project. In addition to the algorithm itself, products, interaction and engineering all need in-depth research and exploration. From algorithm to online service to user experience, from laboratory performance to actual scene system performance, there are still many challenging problems to be solved. Each link has many points that can be cut in. What matters is whether the user's problem is really solved.