A unified face liveness attack detection method and device based on feature dimension reduction and center guidance, equipment and storage medium

Abstract

The application provides a unified face liveness attack detection method and device based on feature dimension reduction and center guidance, a device and a storage medium, wherein a feature extractor trained by multi-class attack supervision is used to obtain high-dimensional features with attack type distinguishability, then the features are processed by dimension reduction to reveal the cluster structure of liveness samples in the feature space, and the geometric center is calculated as the liveness feature center based on the low-dimensional feature set of the liveness samples after outlier filtering in the verification set, and the geometric decision boundary is constructed by the center and the discrimination radius. During detection, only the Euclidean distance between the low-dimensional features of the test sample and the liveness center is calculated and compared with the discrimination radius to complete the judgment. The method converts the detection problem from fitting the attack distribution to modeling the liveness distribution, so that any sample deviating from the liveness center is rejected, thereby achieving effective detection of unknown attacks without relying on prior knowledge of specific attack types.

Images