Face Anti-Spoofing Based On Causal Learning

In recent years,with the explosive data and powerful hardware resources,face recognition technology has developed vigorously.Face anti-spoofing technology is proposed to protect face recognition from physical presentation attacks,including print attack,replay attack and mask attack.It can discriminate the liveness attribute of the presented face and ensures the security of face recognition system.Although existing methods achieve good accuracy in constrained scenarios,they face the unseen attacks,domain shift and privacy leakage problems.This paper explores the use of causal learning for addressing the above issues,which aims to achieves robust and trustworthy face anti-spoofing in realworld scenarios.The research innovations and contributions of this paper are three folds:1.In view of the diversity and difference of attack types in physical media,we formulate the face antispoofing as an open-set recognition problem.Thus,all unknown attacks are defined as open categories.Causal learning based counterfactual synthesis is utilized for abundant samples construction.The extreme value theory is then used to fit the tail data of the known categories,so as to make the decision boundary more precise.The identity-aware contrastive loss is further proposed to help identify and reject the unseen attacks.Experiments show that the proposed approach achieves better performance compared with the mainstream methods,indicating that it can deal with unknown attacks under an open environment.2.To deal with the heterogeneity of liveness attribute in face images,a hierarchical feature modulation approach is proposed to remove interference information from high-level semantics and low-level features respectively.The confounding problem of underlying features is modeled as a distribution mixing problem,and a feature intervention module driven by the expectation maximization algorithm is used to remove confounding factors.It can further extract fine-grained liveness attribute information.Then a local difference histogram feature module is proposed to supplement the orderfree statistical information for the deep structural features.Experiments show that the proposed hierarchical feature modulation method has good accuracy in different scenarios.3.Paying attention to the sensitivity of face anti-spoofing data,we combine the face age estimation for dual causal learning.The perception of high-level semantics such as identity by facial age estimation task is used as an effective guide to add noise to face images.The addition process of noise is guided by the counterfactual attention mechanism and it meets the requirements of differential privacy.Experiments show that the proposed method can protect identity privacy well and do not significantly affecting the accuracy of face anti-spoofing.Moreover,this paradigm can also improve the accuracy of face age estimation.