文献类型：期刊文献

英文题名：Generalization-oriented face forgery detection via discriminative feature analysis and normalization

作者：Li, Xin[1,2];Xu, Bingxin[1,2];Liu, Hongzhe[1,2];Pan, Weiguo[1,2];Xu, Cheng[1,2]

通讯作者：Xu, BX[1];Xu, BX[2]

机构：[1]Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing 100101, Peoples R China;[2]Beijing Union Univ, Coll Robot, Beijing 100101, Peoples R China

第一机构：北京联合大学北京市信息服务工程重点实验室

通讯机构：[1]corresponding author), Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing 100101, Peoples R China;[2]corresponding author), Beijing Union Univ, Coll Robot, Beijing 100101, Peoples R China.|[1141739]北京联合大学机器人学院;[11417]北京联合大学;[11417103]北京联合大学北京市信息服务工程重点实验室;

年份：2025

卷号：31

期号：5

外文期刊名：MULTIMEDIA SYSTEMS

收录：;EI(收录号：20253419045831);Scopus(收录号：2-s2.0-105013893683);WOS:【SCI-EXPANDED(收录号:WOS:001616885100003)】；

语种：英文

外文关键词：Face forgery detection; Generalization study; Texture mining; Feature normalization; Feature disentanglement

摘要：The escalating sophistication of deepfake technologies poses a formidable challenge to the reliability of digital identity verification, necessitating robust face forgery detection systems. Existing methods often struggle with generalization across unseen scenarios, primarily due to their reliance on handcrafted features and a lack of robustness against domain-specific variations. Moreover, the sensitivity of these methods to minor perturbations in forgery methods limits their applicability in real-world settings. To overcome these issues, we introduce a novel framework to boost the generalizability of face forgery detection via discriminative feature analysis and normalization. Specifically, we design a Progressive Texture Mining (PTM) module that leverages center difference convolution and a Global Context module to extract both local fine-grained artifacts and global texture information during the early learning stages. Then, Forgery Aware Normalization (FAN) aligns feature distributions across different forgery styles and classes, promoting intra-class compactness and inter-class separability. Furthermore, Discriminative Feature Disentangle (DFD) addresses the challenge of domain-specific feature sensitivity by selectively disentangling and discarding sensitive channels. Through rigorous evaluation on five public datasets, our method demonstrates superior generalization performance, achieving high accuracy across a spectrum of forgery techniques compared with other state-of-the-art methods.