Exploring the Adversarial Robustness of Face Forgery Detection with Decision-based Black-box Attacks

The emergence of sophisticated face forgery generation technologies has heightened public concerns regarding security and privacy, especially as systems for electronic payments and identity verification increasingly depend on accurate face forgery detection. Despite the success of current detection methods, recent studies indicate that these systems are highly susceptible to adversarial examples, which exploit their vulnerabilities. To address this critical issue, a new study introduces decision-based black-box attacks aimed at improving the adversarial robustness of face forgery detection. The researchers identify challenges in applying existing attack methodologies, such as initialization failures and diminished image quality, and propose novel solutions like cross-task perturbation and frequency decision-based attacks. These innovative approaches leverage the high correlation of facial features across different tasks and the use of frequency cues, respectively. Extensive experiments…