Assessing Backdoor Risk in Deepfake Detection

Jiawen Wang, Boquan Li, Min Yu, Kam-Pui Chow, Jianguo Jiang, Fuqiang Du, Xiang Meng, Weiqing Huang

Abstract: Deepfake images and videos are spreading in cyberspace, threatening individuals and society. Although well-designed deep learning models have achieved satisfactory performance in detecting deepfakes, they face security risks. Compared with deep learning attacks on artificial intelligence systems, backdoor attacks on deepfake detectors have received limited attention.