Self-Supervised Adversarial Training for Robust Face Forgery Detection
Yueying Gao (Communication University of China),* Weiguo Lin (Communication University of China), junfeng xu (Communication University of China), Wanshan Xu (Communication University of China), Peibin Chen (Communication University of China)
The 34th British Machine Vision Conference
Abstract
With the advancement of face forgery technologies, the high-fidelity generation and substitution of human faces have become increasingly prevalent, leading to an emerging research topic of face forgery detection. Despite the outstanding performance of current face forgery detectors in benchmark datasets, their real-life application is fraught with challenges due to complex scenarios. Therefore, we propose a self-supervised adversarial training network to enhance the robustness of face forgery detection, promoting their applicability in real-life scenarios. We generate multiple adversarial examples using a pool of attack strategies and strengthen the sensitivity to perturbations by compelling the model to predict these attack strategies. Additionally, we employ an adversarial training strategy to dynamically generate the most challenging adversarial examples for the current model. A fast training strategy is proposed to reduce the computation cost of adversarial training. Through extensive experiments, we demonstrate that our approach significantly outperforms the baseline and state-of-the-art methods in terms of robustness to perturbations.Video
Citation
@inproceedings{Gao_2023_BMVC,
author = {Yueying Gao and Weiguo Lin and junfeng xu and Wanshan Xu and Peibin Chen},
title = {Self-Supervised Adversarial Training for Robust Face Forgery Detection},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0718.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by