As a newly emerging interactive form, live broadcasting has become one of the most popular industries in recent years. In 2016, it seems that the “war of thousands of regiments” has entered the boom. Happy Hour invested 1 billion yuan in Huya and ME live, and spent 100 million yuan to sign the anchor MISS; Tencent invested 400 million yuan in Douyu, which is valued at $1 billion; The newly established Inke received 80 million RMB investment from Kunlun Wanwei, Fusai and other institutions; Yizhong obtained A round of financing of RMB 60 million; 360 launched pepper, second shot launched a live…. Live network broadcasting is enjoying its best time, live broadcast “tuyere” become the focus of investors.
However, due to the fast pace of network broadcast, the original intention of starting is lost. Anchor behaviors and live broadcast contents become uncontrollable, and vulgar, large-scale and unlimited programs become “explosive”. On December 1, China’s Cyberspace Administration of China (CAC) issued what media described as the “strictest ever” regulations on the management of Live streaming services, in response to the weirdest aspects of live streaming. The regulations require livestream platforms to authenticate users’ real identity information, including mobile phone numbers, in accordance with the principle of “real-name backstage and voluntary reception”. For anchors, the regulation requires more stringent procedures such as checking their identity documents.
In the process of real person authentication, these network live broadcasts need to identify such elements as name, ID number, biological attributes, mobile phone and location. How to accurately judge the real person behind the account while improving user experience is also a major problem plaguing major companies at present. At present, the technology used in aliju security certification can greatly provide automated certification rate and audit efficiency.
Anchor real person certification
Anchor real person authentication, identification OCR based on independent research and development, such as face recognition achieved more than 88% of the automatic authentication rate, and adopt cloud (3 d and duplicate detection, etc.) + side (interactive action, etc.) with the combination of living detection technology to defend the false certification risk of attack, to confirm the identity of the host, reduce the risk of illegal.
Strict real-person authentication is carried out for anchors: the combination of real-person authentication is achieved, that is, according to the authenticated account identity information, the corresponding real natural person can be found accurately.
Real person authentication system schematic diagram
There are many smart technologies involved, two of which are described.
1. Id card OCR
OCR identifies the text in the id card image as the text that can be recognized by the computer, and automatically compares it with the authoritative database such as the public security network to verify the authenticity of the information such as name and number.
Based on the free shooting of certificate images, the process is shown in the following figure. In order to ensure the recognition rate and speed, the traditional algorithm and deep learning algorithm are combined.
Flow chart of id card information detection
The accuracy of OCR algorithm for “name” field is more than 98%, and for “ID number” and “validity period” is 99.5%. The system has strong robustness, and the following cases can be easily identified.
Examples of IDENTIFICATION cards that can be recognized by OCR
2. Biometric recognition
Only face recognition and face in vivo detection are introduced.
Face recognition has surpassed the recognition level of naked eyes in the academic world, but it can be applied on a large scale in practice, because of the complexity of the actual scene and the lack of data. Challenges come from lighting, poses, reshoots, makeup, aging, and poor photo quality.
We compare the user’s personal portrait, id card and head picture in authoritative database to verify the authenticity of identity. The algorithm can make the automatic pass rate of legitimate users reach 93% under the condition of 0.1% acceptance rate.
Face recognition includes image acquisition, face detection, in vivo detection, key point location, feature extraction, recognition engine and other modules.
Face recognition system
2.1 Face detection
Boosting+RCNN framework was adopted.
Image of face detection results
2.2 In vivo detection
The purpose of living detection is to ensure that the users to be authenticated are “living people”, rather than face photos and videos taken in advance or remade, so as to prevent false authentication and reduce the risk of illegal anchors.
Example diagram of living detection
In vivo detection module includes:
Face detection
Detect whether there is a face, and can not be multiple faces, to prevent the switch between different people or people and photos.
3 d detection
Verify whether it is a stereo portrait to prevent flat photo or video attacks.
3D detection diagram
Living algorithm detection
Verify that the user’s operation is normal and assign the user to perform random actions (stare, shake head, nod, blink, move the phone up and down, etc.).
Continuity detection
Prevent switchover halfway.
Duplicate detection
Deep learning technology is used to distinguish whether the portrait obtained is a reproduction of the screen and photos.
2.3 Face key point positioning
Locate eyebrows, eyes, nose, mouth, etc. The main methods are: the method based on parameter model; Regression based method; Methods based on deep learning.
We use feature-based regression method and deep learning method to train the key point locating model.
Schematic diagram of human face key point detection (Picture material source network)
2.4 Feature Extraction
Schematic diagram of Maxout structure
We adopt both traditional face features (WLD, HOG, LBP, Gabor, etc.) and features based on deep learning (fusion network based on VGG, GoogleNet and Maxout) for face recognition.
2.5 Identification Engine
For traditional face features, we use SVM to measure Pairwise distance. DL face features, fine tuning of the learned classification model.
Real person authentication automation is based on the recognition results of face and OCR, based on big data, comprehensive use of user behavior characteristics, multi-dimensional information fusion, finally get a comprehensive decision model, realize the process of automatic decision. And in the development of the Internet in a variety of business, real person authentication technology application scenarios are complex, the requirements of technical indicators are not the same, Aliju security to provide real person authentication technology to biometric, wireless security technology as the support, to ensure the effectiveness of real person authentication.
At present, Alibaba’s face recognition technology has been applied in a large scale in practical scenes. In actual combat, relevant performance indicators are FPR (False Positive Rate) 0.1%, TPR (True Positive Rate) 96%, the recognition accuracy is far higher than human eye recognition. Ali gather security real person authentication technology more big data real-time risk management as the core, can be real-time judge each user authentication motivation, for users with different levels of risk taking different ways of certification, to ensure normal users can easily and fast to disclose the information, and users can’t risk simply by theft information through the authentication, to ensure the authenticity of certification.
Previous review: AI power in risk prevention and control of Alibaba live broadcast content
For more security technology articles from Ali, please visit Aliju security blog