Joint Statistical and Causal Feature Modulated Face Anti-Spoofing
2023 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO, ICME(2023)
摘要
In this paper, we propose a hierarchical feature modulation (HFM) approach for stable face anti-spoofing in unseen domains and unseen attacks. The conventional multidomain based generalizable approaches likely lead to local optima due to the complicated or heuristic learning paradigm. Inspired by the fact that high-level semantic disturbances and low-level miscellaneous bias jointly cause the distribution shift, HFM aims to modulate the fine-grained feature in a hierarchical manner. Specifically, we complement the structural feature with patch-wise learnable statistical information, i.e. local difference histogram, to relieve the overfitting on high-level semantics. We further introduce the structural causal model (SCM) with imaging color model to reveal that presenting mediums and capturing devices destroy the liveness-relevant information from the low level. Thus we model this hidden entanglement as a distribution mixture problem and propose the expectation-maximization (EM) based causal intervention to remove these miscellanies. Experimental results on public datasets demonstrate the effectiveness of HFM, especially in out-of-distribution settings.
更多查看译文
关键词
Face anti-spoofing,Statistical feature,Causal intervention,Expectation maximization
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要