Behind the success of helping the rich find their relatives: face recognition technology is still in the "deep water area" in the era of large models
Every reporter: Ke Yang, Wang Jing Every editor: Yang Xia
Face recognition is one of the earliest and most mature technologies in the field of AI. Nowadays, "face brushing" has become a norm in various occasions, taking transportation scenarios as an example, face recognition has been widely used in boarding, gate, security check and other links.
In addition to the scene with the nature of audit and inspection, recently, the news that a wealthy man in Hebei has recovered his son after 25 years has attracted public attention, and Geling Shentong (688207. SH, share price 20.94 yuan, market value 5.4 billion yuan) was also pushed to the forefront. On the afternoon of December 4, Geling Shentong issued a statement saying that for the search for relatives, the company mainly cooperates with the police to provide tools such as technology and algorithms, and then hands them over to the police for application.
In addition to Geling Deep Pupil, there are also many listed companies in the A-share market that are also engaged in the research and application of face recognition, including Yuncong Technology-UW (688327 .SH, stock price 17.25 yuan, market value 17.9 billion yuan), iFLYTEK (002230. SZ, share price 45.38 yuan, market value 105.086 billion yuan), Hanwang Technology (002362. SZ, share price of 22.8 yuan, market value of 5.574 billion yuan). At the same time, with the rise of large model technology, many companies choose to launch visual large models this year, and based on large model training, face recognition technology has also ushered in innovative opportunities.
However, as a new technology for recording biometric data, face recognition technology cannot be ignored while "technology for good" and serving human beings.
Image source: Visual China
Facial recognition algorithms help find relatives
Recently, the news of "Hebei billionaire Xie Kefeng's 25-year-old dream of his abducted son Xie Qingshuai has attracted much attention. Behind the reunion of relatives, it is the power of science and technology that is playing a role, Geling Shentong said on the official Weibo that the company's independent research and development of the "cross-age and related face comparison algorithm" is indispensable.
According to public information, the main business of Geling Shentong is to deeply integrate computer vision technology, big data analysis technology, robotics technology and human-computer interaction technology with application scenarios, and provide artificial intelligence products and solutions for smart finance, urban management, commercial retail, rail transit operation and maintenance, sports and health, and metaverse. In 2022, the company landed on the Science and Technology Innovation Board.
According to its official WeChat account, the police often face three major problems when searching for abducted children: the incomplete original data leads to a small number of valid clues, the large amount of data brought by cross-region, and the large variability caused by cross-age. In fact, with the passage of more than a decade or decades, the appearance of abducted children has already undergone earth-shaking changes, from young children to teenagers and even middle age.
In order to help these parents find their children as soon as possible, the police have done a lot of work: manually checking and matching them one by one, repeatedly searching and comparing them according to the relevant information of the abducted children, and conducting multi-dimensional analysis and de-elimination of suspected objects, investing a lot of time and manpower.
How to deal with complex data correctly and quickly can help the police narrow down the scope of suspicion and test the technical capabilities of enterprises. In this regard, Geling Shentong independently developed the "cross-age and related face comparison algorithm", and has helped the police find 4 abducted children.
In terms of specific operation methods, Geling Shentong detailed on the official public account, "Screening out thousands of pictures from tens of thousands of local picture information, and then comparing these thousands of pictures with the information of immediate relatives and family members, based on genetic relationship, the similarity of face features between relatives will be relatively high, using this law, the cross-age face recognition algorithm provided by Geling Shentong will screen out the suspected ones with high relevance after recognition, assign and rank, narrow the search range, and finally take DNA identification as the final judgment." After repeated testing and tuning, the results were also surprising."
Regarding the technical advantages of Geling Shentong's "cross-age same-related face comparison algorithm", the company said: "After ten years of practical application of the algorithm, the company has built face recognition models, which can capture the subtle features and similarities of people, so that they can still perform well in complex cross-age recognition scenarios. Secondly, the model is trained by a large number of samples, and the samples of different ages are used to generate face images of different people and different ages, so as to improve the accuracy of cross-age recognition. Thirdly, an efficient distributed training algorithm is used to train a large face recognition model, which has stronger generalization ability for cross-age recognition, so that it can adapt to a larger span of age recognition. ”
In terms of performance, Geling Shentong's third quarter report for 2023 shows that the company's operating income in the first three quarters was 225 million yuan, a year-on-year increase of 14.76%, and the net profit attributable to shareholders of listed companies was -17.279 million yuan, a year-on-year decrease of 8.0%. In terms of the third quarter, the company achieved operating income of 67.155 million yuan, a year-on-year decrease of 14.48%, and net profit attributable to shareholders of listed companies -19.394 million yuan, a year-on-year decrease of 464.24%.
As for the reasons for the decline in performance in the third quarter, Geling Shentong said in the financial report that it was mainly due to the year-on-year decrease in operating income and the increase in R&D investment.
The era of large-scale models has once again ushered in technological innovation
Technology is not an instantaneous innovation.
Prior to this, face recognition technology has been the first technology to be applied and implemented in artificial intelligence technology, and it is also the most direct intersection between AI and the real world. In the field of artificial intelligence, which has been trapped in the application for a long time, face recognition, as one of the first technologies to land, has given birth to a number of technologies based on Hikvision (002415. SZ, share price of 33.6 yuan, market value of 313.508 billion yuan), iFLYTEK and other star enterprises.
The 22 A-share face recognition concept stocks screened by Wind cover two giants with a market value of 100 billion yuan, namely Hikvision and iFLYTEK, and there are also seven companies with a market value of more than 10 billion.
In terms of performance, in the third quarter of 2023, there are five companies with revenues of 10 billion yuan, of which Hikvision achieved operating income of 61.275 billion yuan in the first three quarters, ranking first, Dahua shares (002236.SZ, stock price 18.9 yuan, market value 62.265 billion yuan) and iFLYTEK achieved revenue of 22.278 billion yuan and 12.614 billion yuan respectively in the same period.
Advanced deep learning algorithms, high-resolution cameras, and powerful computing power have made a qualitative leap in accuracy and speed of face recognition systems. From the early days of biometric tools to today's widespread use in security, finance, retail and other fields, face recognition technology has undergone earth-shaking changes.
Image source: Visual China
However, in recent years, as the application scenarios of face recognition have become stable, the expansion of new scenarios has also progressed slowly.
As the technology matures, how can companies expand into new application scenarios?
According to the interview and research of the reporter of "Daily Economic News" from December 5th to 6th, the current expansion direction of the industry can be summarized into two aspects: on the one hand, it is the expansion and technological innovation around the vertical industry, and on the other hand, it is the multimodal expansion in the context of large model technology.
Xu Fangming, marketing director of Hanwang Zhiyuan Marketing Center of Hanwang Technology, said in a written interview with the reporter of "Daily Economic News" that on the one hand, Hanwang Technology has launched comprehensive solutions suitable for vertical industries to meet the needs of scenarios in an all-round way, such as launching comprehensive solutions for park management, and on the other hand, it has integrated with other innovative technologies to meet the richer needs of customers in different scenarios, such as launching identity authentication based on face + gait recognition to meet higher accuracy.
Yuncong Technology told the "Daily Economic News" reporter that if the vision of artificial intelligence is to build a robot with a brain, nerves, torso and limbs, then the hardware foundation is the torso and limbs, and more importantly, to make the robot can see and think, hear and speak, it is necessary to build nerves and brains. Therefore, face recognition is only a branch technology of machine vision, and Cloudwalk Technology is committed to developing a closed loop of language, vision, and multi-modal large model technology, so that users can get a more complete application experience. Based on the large model of Yuncong Calm, Yunyue Technology released the AI elf (AI-Agent) Yunyue, which not only has a high degree of anthropomorphic presentation, but also is closer to the level of real people in terms of image, action and intelligence, and can understand, understand, remember, self-learn, and interact naturally with people.
Under the boom of large models, it has become an industry consensus that Cloudwalk Technology also chooses to rely on large model technology to seek expansion. Since the beginning of this year, a number of companies have announced the launch of visual large models, and the continuous improvement of the performance of visual large models has revolutionized the accuracy and robustness of face recognition technology. Geling Shentong mentioned in a previous statement that the company uses efficient distributed training algorithms to train large face recognition models, so that it has stronger generalization capabilities for cross-age recognition and can adapt to a larger span of age recognition.
The first to land is stuck in application security
On the back of the acceleration of technological innovation by large models, the security of face recognition is also ushering in new challenges.
"Compared with traditional static live attacks, the new AIGC-based and large-model deepfake attacks iterate faster, which means that the defense software for face recognition needs to be updated and upgraded more frequently. High-intensity offense and defense around face recognition will become the new normal in the industry. Xiao Zihao, co-founder and algorithm scientist of RealAI, said in a written interview with a reporter from the "Daily Economic News".
At present, face recognition has become a common way of artificial intelligence application in people's production and life, and applications such as face brushing to enter the door, face brushing to pay, and face punching card have quietly changed people's living habits and improved people's production efficiency. According to Statista, a data and statistics company, the global face recognition market size was about $5.01 billion in 2021 and is expected to double to $12.92 billion in 2027. With the in-depth application of face recognition technology in various industries, the scale of the Chinese face recognition market is expected to further develop.
However, face recognition technology has also exposed many privacy and security issues in its application, which has attracted social attention. For example, smart express lockers have been exposed to be able to easily break through face recognition with a single printed photo, the "3.15" party exposed that real estate companies process facial information without the user's perception and authorization, and Mate compensated 1.6 million users for abusing the facial recognition function with a total of 650 million US dollars.
Image source: Visual China
Xiao Zihao also mentioned in an interview that there is a risk of leakage of face data, and there is a precedent for it. Failure to protect the data collector will lead to large-scale leakage, causing great harm to the personal and property safety of individuals, and may even threaten public safety. At the same time, the technical shield of the facial recognition system is not unbreakable. With the rapid development of artificial intelligence, especially generative artificial intelligence, "hackers" are constantly "iteratively upgrading", including adversarial sample attacks, deep fake attacks, etc., which continue to challenge the security defense line of face recognition systems. Therefore, countermeasure technology will definitely become an important direction in the field of artificial intelligence security.
In terms of ensuring the security of facial recognition applications, the industry generally believes that it is necessary for the government, industry, academia and research institutes to work together to improve the industry's security compliance capabilities. A few days ago, the national standard "Test Method for Information Technology Biometric Recognition Face Recognition System" (GB/T 42981-2023) proposed and managed by the National Information Technology Standardization Technical Committee (SAC/TC 28) was released and will be officially implemented on April 1, 2024.
This standard specifies the general requirements for the testing of facial recognition systems, and describes the functional test methods, performance test methods, and liveness detection test methods of facial recognition systems. It is applicable to third-party inspection and testing institutions to carry out face recognition system testing, and relevant institutions that develop and apply face recognition systems refer to carry out testing activities.
"The implementation of the new national standard will provide a more standardized and unified standard for the application of biometric identification technology, and further promote the development and popularization of the technology. Yuncong told reporters from the technology aspect.
Xu Fangming of Hanwang Technology said that the formal implementation of national standards will first help standardize the entire face recognition industry, corresponding to the entire industry, which can improve the product quality of the entire industry, promote the development of standardization, enhance market competition, and enhance consumers' trust in face recognition technology; Finally, companies can offer solutions that meet the new standards to customers to meet market needs and expand market share.
In terms of ensuring the security of face recognition applications, Xiao Zihao suggested that companies should strengthen the investigation of hidden risks of face recognition. In particular, applications related to property, law, and government affairs need to pay attention to some security risks generated by face recognition, but there is a delay effect in the risk explosion. In view of the problems identified, it is necessary to design a new risk control mechanism to plug the leakage. Even after plugging the leak, it is still recommended that the company regularly conduct attack and defense drills on current risks such as face data privacy protection and face recognition software vulnerabilities, and upgrade the face firewall system frequently to ensure effective defense against new attacks.
(Cover picture source: Daily Economic News data map)
National Business Daily