Consolidate privacy and security, and create safe and reliable payment-level live experience detection capabilities

Content source: Huawei Developer Conference 2021 HMS Core 6 AI Technology Forum, the keynote speech "Consolidate Privacy and Security-Machine Learning Services Create Safe and Reliable Payment-Level Live Experience Detection Capabilities."

Speaker: Ban Guangwei, Huawei Machine Learning Service Product Architect

Hello everyone! It is an honor to introduce to you the payment-level live detection capabilities of Huawei's machine learning service.

In daily life, living body testing services are widely used, such as real-name authentication and mobile phone unlocking. Compared with the former, the concept of face recognition is actually more well-known, and it is used in scenarios such as high-speed train brakes identity verification, face verification in exhibition areas, security and monitoring systems. Public places perform specific identification based on facial features, and people usually unlock private devices. These have gradually become social rigid needs.

But the face recognition algorithm itself is the extraction of face information. As for whether the extracted information is a real person, its discriminating ability is not high. For example, using high-simulation pictures, precision plaster or 3D modeling masks can break many face recognition algorithms. The emergence of live detection technology just made up for this shortcoming, and greatly improved the security of face recognition.

The value and challenge of in vivo testing

With the increasing popularity of "face-brushing" scenes, the value of in vivo detection has become more prominent. It is widely used in financial payment, medical government affairs, insurance and financial management and other fields. At present, the daily call volume of face authentication of Huawei's machine learning service is relatively high, indicating that developers are very interested in this technology.

Of course, live detection also faces many challenges. The abundance of application scenarios brings some uncertainties. For example, different application scenarios have different requirements for live detection performance; the diversity of equipment makes the performance difference of high, medium and low equipment large, ethnic diversity, and diverse environmental changes. There are also various forms of attacks that hinder live detection, such as static screen attacks, static video attacks, print photo attacks, photo burrowing, 3D mask model attacks, photo activation attacks, ROM injection attacks, script attacks, interface attacks, IP/phone Attacks, group control equipment attacks, etc.

There are three common types of live detection in the industry, all of which can be screen-based, paper-based, and mask-based, and the cost is from low to high: one is RGB live detection , using RGB camera; two is near infrared Living body detection uses infrared cameras to identify the infrared characteristics of living beings; Third is 3D living body detection , which uses structured light/TOF depth cameras to directly recognize the 3D structure of human faces or human bodies with higher security.

Three live detection solutions for machine learning services

Huawei's machine learning service provides developers with three live detection solutions, and optimizes algorithm performance to ensure a good application experience——

• Silent Living Detection Program . This can prevent the mobile phone from being unlocked when the party is not present, ensuring safety. Living body data covers scenes such as lighting, facial accessories, gender, hairstyle, mask material, etc. The model design adopts the lightweight convolution module; the model deployment adopts the Mindspore-lite reasoning framework to tailor the operator to achieve the ultimate packet size.

• Interactive live detection program . It is suitable for scenarios that require human-computer interaction, such as banking, finance, and medical care. According to the instructions, the person involved blinked, opened his mouth, shook his head left, shook his head right, and looked at five kinds of actions, and then randomly selected three to make the detection more secure and make the fake face disappear immediately. At the same time, it supports guided detection, such as "prompts for faces that are too close or far away; prompts for dark light; prompts for mask occlusion", etc., making the interaction more friendly.

  

• colorful living body is a new technology, without the user's cooperation , stay for 2-3 seconds. It uses three primary colors (red, green, blue) and yellow as a specific light source for encoding, so that the light source device emits a specific coded light, and the object can be reflected according to the specific light, and multiple frames of two-dimensional human faces are mapped into three-dimensional space. According to the changes between frames, more accurate depth information estimation is performed. It is characterized by high accuracy and can prevent mask attacks, video attacks and printing attacks. It is suitable for application scenarios such as payment, authentication, unlocking, and child mode.

  

Machine learning service live detection, open architecture

Currently, the silent live detection capability has been launched on the official website for developers to use, and the interactive live detection capability will also be launched in the near future. This is all included in the open architecture of Huawei's machine learning service live detection.

• application layer suitable for mobile phone unlocking and real-name authentication scenarios.
• connection layer provides developers with two ways to integrate: fullSDk supports end-to-end integration of all content without additional networking or download; IiteSDK provides a lightweight interface package, which can be packaged into your own application with only tens of K. You can download it directly from the Huawei App Market.
• hardware layer supports ordinary USB cameras, and there is no special requirement for the resolution of the mobile phone.
• system layer supports the compatible live detection capabilities of the Android system and the Hongmeng dual framework.
• algorithm layer implements two ways of silent living and interactive living. There are many breakthroughs in algorithms for silent living detection. We cooperated with data companies to collect more than 200 types of data scenarios to ensure coverage of the diversity of user usage scenarios. Its training data has reached tens of millions of levels. The interactive living body provides developers with a complete set of guidance controls and actual algorithm calling framework. Every developer can refer to the interactive UI for simple integration. Currently, five random actions are supported, and we will open more actions in the future for everyone to integrate and choose.

In the future, we plan to provide safer silent live detection capabilities to reach the payment level security level. Added multi-modal live detection capability for line-of-sight estimation. The user completes the screen prompts (such as looking at the blue circle, the largest number of eyes, etc.). The application distinguishes live and non-living bodies by capturing the movement direction and gaze direction of the eyes. The binocular live detection capability is also under technical planning.

more details>>

Visit Huawei Developer Alliance official website
Obtain development guide document
Huawei mobile service open source warehouse address: GitHub , Gitee

and learn about the latest technical information of HMS Core for the first time~