Covered Style Mining via Generative Adversarial Networks for Face Anti-spoofing. (December 2022)

• Record Type: Journal Article
• Authors: Wu, Yiqiang; Tao, Dapeng; Luo, Yong; Cheng, Jun; Li, Xuelong
• Is Part Of: Pattern recognition. Volume 132(2022)
• Publication Date: 2022-12
• Subjects: Face anti-spoofing -- Generative adversarial networks -- Deep learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108957
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel framework, CSM-GAN, is proposed to achieve face anti-spoofing using style transfer technology. CSM-GAN converts the original binary classification task into a style transfer task and, unlike existing methods, CSM-GAN mines potential difference distributions without introducing prior information and generates a difference map. To achieve end-to-end training and style transfer, an updatable three-channel difference map is designed to combine each sub-module, which provides richer photography style information than traditional fixed one-dimensional vectors. To prove the effectiveness of proposed method, extensive experiments are conducted on published face anti-spoofing datasets, demonstrating its superior performance to current stateof- the-art. Abstract: Face anti-spoofing, a biometric authentication method, is a central part of automatic face recognition. Recently, two sets of approaches have performed particularly well against presentation attacks: 1) pixel-wise supervision-based methods, which intend to provide fine-grained pixel information to learn specific auxiliary maps; and 2) anomaly detection-based methods, which regard face anti-spoofing as an open-set training task and learn spoof detectors using only bona fide data, where the detectors are shown to generalize well to unknown attacks. However, these approaches depend on handcrafted prior information to control the generation of intermediate difference maps and easily fall into local optima. In this paper, we propose a novel frame-level face anti-spoofing method, Covered Style Mining-GAN (CSM-GAN), which converts face anti-spoofing detection into a style transfer process without any prior information. Specifically, CSM-GAN has four main components: the Covered Style Encoder (CSE), responsible for mining the difference map containing the photography style and discriminative clues; the Auxiliary Style Classifier (ASC), consisting of several stacked Difference Capture Blocks (DCB) responsible for distinguishing bona fide faces from spoofing faces; and the Style Transfer Generator (STG) and Style Adversarial Discriminator (SAD), which form generative adversarial networks to achieve style transfer. Comprehensive experiments on several benchmark datasets show that the proposed method not only outperforms current state-of-the-art but also produces better visual diversity in difference maps.