当前位置:网站首页>Research and application of face security technology

Research and application of face security technology

2020-12-07 19:20:46 Youtu Laboratory

In recent years , Face recognition technology has been widely used in real life , Device unlock 、 Brush face payment and entrance guard security can be seen everywhere , And face editing and other new technologies are also emerging , At the same time, it also causes people to worry about face safety .

In order to let you have a deeper understanding of face security technology ,11 month 11 Japan , Tencent Youtu held youyou Tech Online Salon , Laboratory researchers are surrounded by 《 Research and application of face security technology 》 Live sharing , The form of face attack and the corresponding defense algorithm are introduced , The prospect of face security is also discussed .

Before we talk about face safety , Let's first learn about face attack methods ?

There are three main forms at present : The first physical attack on the face , Based on paper 、 The screen 、 Face attack by plaster and other media . The second type of face content generation attack , Based on the generation of countermeasures network 、3D Modeling, etc 、 Edit faces . The third kind of face confrontation attack , Based on the method of counter attack, the sample is generated , Make the deep learning model output wrong results , There are mainly two types of attacks , Including digital image counter attack and physical form counter attack .

If there are criminals using the above form of face attack , It will do harm to face safety : One is the wrong authentication ; The second is the emergence of false news .

To ensure face safety , In view of the above face attack form, Tencent excellent graph has developed the corresponding defense algorithm .

For physical media attacks on human faces , Tencent Youtu has developed a face detection technology , This technology can be used to determine whether the object is a real person in the authentication scene , Defense photos 、 Mask 、 Screen remake and other types of attacks .

The research of living body detection is earlier in academic circles , From traditional manual features to deep learning , From binary constraints to multiple auxiliary constraints . Tencent excellent graph has complete single 、 Multimodal living detection method , There are three main directions of single mode detection in vivo : Firstly, the essential characteristics of living body are explored through feature decoupling , Improve the generalization of tasks . secondly , From a data point of view , Using self-monitoring constraints to further remove the independent features in vivo , Improve generalization .

Last , To further enhance security , Tencent Youtu has designed a new in vivo detection method - Light in vivo detection . There are two characteristics of light in vivo detection : The first is to rely on hardware , It doesn't rely on extra camera hardware , Just have a screen that can emit different color sequences , We can complete the in vivo verification . The second is captcha strategy , Combined with cryptography strategy , Through screen colors 、 The brightness is random , To improve the security of the system .

For content composition attacks , Tencent Youtu has carried on the research in the field of face forensics .

Face forensics is aimed at face generation 、 Face editing for identification , There are four common types of face forgery : Face exchange 、ID In exchange for 、 Attribute editing and global face generation . Regarding this , Tencent optimal graph improves the accuracy of face forensics based on multiple auxiliary constraints , And focus on the texture changes caused by face editing , As well as the boundary noise introduced by face fusion , Data preprocessing and multi branch network structure are designed , In order to finally combine the multi-scale image features to do the final true and false discrimination .

Tencent Youtu participated in a large-scale fake face detection competition this year Depfakes Detection Challnge, There are as many teams as 2265 the , The main task is to detect whether the face or sound in the video has been tampered with . The game is very challenging , Tencent Youtu starts from video data processing and network structure design , Using attention mechanism to improve the accuracy of forensics . In the final result of the game , The algorithm of optimal graph has reached the top of the public list 1%, Gold medal on the private list .

Through this competition Tencent excellent graph summed up three points about face forensics experience , One is that the data enhancement of partial removal and data mixing is more effective ; Second, in the network structure design , Try multi-scale 、 Fusion of multiple models , And introduce sequence analysis model and attention mechanism ; Third, the current forensic methods lack of methods based on exploring sensor noise and image synthesis process .

Face to face attack , Tencent Youtu has studied both attack and defense .

At present, there are three kinds of attack forms in the face physical confrontation attack industry :ADV – hat、ADV - nose as well as ADV- glass, These three forms are relatively obvious visually . In order to further enhance the concealment , Tencent Youtu has designed a covert makeup attack , Make use of the generated countermeasures network , Part of the makeup of anti noise camouflage adult eye shadow .

Besides , It also studies the face defense . There are two main research directions : The first is to counter sample testing , The input image is detected to be a counter sample , And refuse to recognize ; The second is recognition model reinforcement , Even if the input image is a counter sample , Recognition models can also make correct judgments .

The above mentioned face security technology , Tencent Youtu has been successfully applied .

Youtu face core technology has opened an account in finance 、 Intelligent government and online supervision have been widely used in many scenarios , Such as : Remote account opening business of Weizhong bank 、 Taiking bank - Local deployment EKYC 、 China Unicom - Tencent big and small king card remote card business 、 Didi driver remote authentication service .

Offline , The multimodal living detection scheme of the optimal graph is paid in the face 、 Travel and authentication multiple scenarios have been widely used , And I got it BCTC Enhanced biopsy Certification . in addition , The application of face anti-counterfeiting technology has also effectively cracked down on pornographic illegal production 、 Politicians make fun of 、 False evidence, etc .

Taiping said , With the evolution of face attack technology and the increase of attack types , The future scene application of face technology will face many challenges . Tencent Youtu will further improve the model generalization based on the method of domain generalization 、 Based on the method of generating network, the explainability of network learning features is explored, and the essential characteristics of each task are explored based on decoupling learning , And continue to conduct in-depth research in these directions .

Today, , From the era of big data to “ The era of face painting ”, Face technology needs to build a strong defense line of information security . As one of Tencent's top AI labs , Tencent Youtu has always focused on computer vision , Focus on face recognition 、 Image recognition 、OCR、 machine learning 、 Data mining and other fields to carry out technical research and development and industry landing .

️ Click on 「 Read the original 」, You can go back to this wonderful live sharing .

Original statement , This article is authorized by the author + Community publication , Unauthorized , Shall not be reproduced .

If there is any infringement , Please contact the yunjia_community@tencent.com Delete .

版权声明
本文为[Youtu Laboratory]所创,转载请带上原文链接,感谢
https://chowdera.com/2020/11/20201114235945157e.html