Black-Box Attacks on Interpretable Deepfake Detectors

This paper highlights the growing concerns over the misuse of deepfake technology by meticulously exploring the vulnerability of interpretable deepfake detectors to adversarial attacks. Our study uses a systematic framework consisting of four key stages: identifying a target model, training a surrogate model, executing attacks on the surrogate model, and validating the attack on the target model. Utilizing deepfake detection models, the study systematically evaluates their susceptibility to adversarial manipulation using cutting-edge attack strategies such as Sign-OPT, Hop-Skip-Jump, Square, Parsimonious, Fast Gradient Sign Method, and Projected Gradient Descent. The efficacy of these attacks is assessed through classification accuracy. Notably, findings underscore the effectiveness of these attacks, particularly when applied to fake datasets. Furthermore, the paper tests the interpretability of attacks on existing deepfake image detector technologies, using Euclidean distance (L2), and Multi-Scale Structural Similarity Index Measure (MS-SSIM) metrics. The results show an average increase of 0.08 in the MS-SIM metric proving the enhanced quality of adversarial examples generated. Through empirical analysis, the paper emphasizes the critical importance of achieving low L2 distance and high MS-SSIM values in designing effective adversarial attacks, contributing to the advancement of defense mechanisms against evolving adversarial threats in multimedia technology.