Face Liveness Detection Using Selected Image Pairs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face liveness detection technologies either expose user interactions, making them vulnerable to attacks, or struggle with collecting face images with significant changes, leading to unreliable detection results.
Innovation Solution
A method that continuously captures face images, selects pairs with large attribute differences, and inputs them into pre-trained models for detection, hiding user interaction information and improving detection accuracy and security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If user interaction information is exposed in the detection process, then the detection can verify liveness, but the security of the detection algorithm deteriorates due to vulnerability to attacks
Solution Approach 1:
The patent extracts and removes user interaction information from the detection process. Instead of requiring visible user actions (eye blinking, head movement), the system selectively captures only the necessary facial region changes caused by these actions, discarding the interaction metadata that would expose the detection methodology to attackers.
Solution Approach 2:
The patent introduces an intermediary mechanism - a detection model that processes facial images without exposing the interaction logic. The model acts as a mediator between the user's natural actions and the liveness verification, transforming interaction-based verification into a passive image analysis process that hides the detection algorithm's vulnerabilities.
2Measurement precision
If face images with significant attribute differences are collected, then detection accuracy improves, but the complexity of collecting and processing such images increases
Solution Approach 1:
The patent applies preliminary action by pre-training detection models on diverse facial images with various attribute differences (lighting, pose, expression) before deployment. This pre-processing of training data eliminates the need for complex real-time image collection and processing during actual detection, as the model is already prepared to handle diverse image variations.
Solution Approach 2:
The patent leverages parameter changes in facial images (lighting conditions, head pose angles, facial expressions) as natural variations to improve detection accuracy. Instead of actively controlling for these changes during image collection, the system utilizes them as inherent diversity in the input data, simplifying the collection process while maintaining high detection precision.
Data Source
AI summary
Some embodiments of this specification provide face liveness detection computer-implemented methods, devices, and media. In an example, multiple face images continuously captured for a first user are obtained. Two face images are selected from the multiple face images for a target attribute in one or more predetermined image attributes, where a degree of difference between two attribute values of the target attribute corresponding to the two face images satisfies a predetermined criterion. Both of the two face images are input into a pre-trained face liveness detection model corresponding to the target attribute to obtain a detection result corresponding to the target attribute.