A Robust Anti-Spoof Identity Verification Framework Using Foundation Vision Models and Behavioral Liveness Reasoning

Abstract

Face recognition technologies are progressively being faced with the challenge of sophisticated presentation attacks like the use of replay videos, 3D masks, and real-time deepfakes. Liveness detection methods that were relying on blinks and head movements are not sufficient anymore, as contemporary generative AI can very realistically imitate human facial behaviors. In this paper, the authors present a hybrid anti-spoofing identity verification framework that integrates behavioral liveness reasoning and Foundation Vision Models (FVMs). FVM extracts very detailed visual artifacts with respect to texture, lighting, and material inconsistencies. Temporal liveness module is dedicated to the analysis of natural facial dynamics, for instance, micro-movements, blinking patterns, and subtle head motion. A fusion layer serving different modalities is in charge of the integration of spatial and temporal cues leading to the output of a very reliable authenticity decision. The framework was put to test on publicly available face presentation attack datasets and it showed very good performance when it came to the task of telling apart real faces from the spoofed and deepfake ones. The experimental results signified that the suggested method provided better endurance than in the case of appearance-only techniques while still being appropriate for real-time implementation. The system is compatible with mobile devices, kiosks, and cloud-based authentication platforms thus it can very well be used in a variety of applications such as banking, e-KYC, and secure access control.