Towards generalizable face forgery detection via mitigating spurious correlation
Neural Networks · 2025
DOI: 10.1016/j.neunet.2024.106909 ↗Kurz gefasst
Fachartikel in der Zeitschrift Neural Networks, 2025. Verfasst von Ningning Bai und 5 weiteren Autor:innen. Beteiligte Einrichtung: Department of Mathematics, Xi'an University of Technology, Xi'an 710048, China. Electronic address: ningnbai@163.com. Hauptthemen laut MEDLINE: Deep Learning. Die Arbeit wurde bisher 7-mal zitiert (Crossref) und nennt 91 Literaturangaben.
Abstract (Originalsprache)
The continuous advancement of face forgery techniques has caused a series of trust crises, posing a significant menace to information security and personal privacy. In response, deep learning is being employed to develop effective detection methods to identify deepfake images and videos. Currently, most detection methods generally achieve satisfactory performance in intra-domain detection. However, their effectiveness remains uncertain when detecting unknown forgeries and cross-domain scenarios. From a causal learning perspective, our analysis reveals that the failure of detectors in cross-domain detection primarily arises from spurious correlations between irrelevant features and category labels (real and fake). In this study, we propose the Feature Independence Constrainer (FIC) to alleviate spurious correlations and enhance the generalization for detection method. Specifically, FIC maps features to a Reproducing Kernel Hilbert Space and calculates the covariance matrix among the kernel-mapped features. Dynamic weight parameters are incorporated into the covariance matrix for iterative optimization. Consequently, statistical correlations between features are eliminated, ensuring independence between each pair of features and mitigating the impact of spurious correlations. Additionally, we introduce fine-grained high-frequency components to force the detection model to learn comprehensive forgery-related artifacts while avoiding spurious correlations between irrelevant features and labels. Furthermore, a Feature Alignment Module (FAM) is designed to explore higher-order dependencies between the spatial and frequency domains, allowing for the extraction of richer potential forgery traces. Quantitative and qualitative experiments demonstrate that the proposed method achieves competitive performance across multiple face forgery benchmark datasets, outperforming state-of-the-art methods.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Mathematics, Xi'an University of Technology, Xi'an 710048, China. Electronic address: ningnbai@163.com.
- Department of Mathematics, Xi'an University of Technology, Xi'an 710048, China. Electronic address: xfwang66@sina.com.cn.
- School of Computer Science and Engineering, Xi'an University of Technology, Xi'an 710048, China. Electronic address: rdhan1988@126.com.
- Department of Mathematics, Xi'an University of Technology, Xi'an 710048, China. Electronic address: houjianpeng55@126.com.
- School of Computer Science and Engineering, Xi'an University of Technology, Xi'an 710048, China. Electronic address: q_wang96@126.com.
- School of Software Engineering, Xi'an Jiaotong University, Xi'an 710048, China. Electronic address: pangsm@xjtu.edu.cn.
Zitieren
APA
Bai, N., Wang, X., Han, R., Hou, J., Wang, Q., & Pang, S. (2025). Towards generalizable face forgery detection via mitigating spurious correlation. Neural Networks. https://doi.org/10.1016/j.neunet.2024.106909
Vancouver
Bai N, Wang X, Han R, Hou J, Wang Q, Pang S. Towards generalizable face forgery detection via mitigating spurious correlation. Neural Networks. 2025. doi:10.1016/j.neunet.2024.106909
BibTeX
@article{Bai2025,
author = {Bai, Ningning and Wang, Xiaofeng and Han, Ruidong and Hou, Jianpeng and Wang, Qin and Pang, Shanmin},
title = {Towards generalizable face forgery detection via mitigating spurious correlation},
journal = {Neural Networks},
year = {2025},
doi = {10.1016/j.neunet.2024.106909}
}
EUVIMED