Exploiting Cross-Modal Coherence Shortcuts: Adversarial Attacks on Audio-Visual Deepfake Detection
Abstract
Audio-visual (AV) deepfake detectors have become increasingly effective by exploiting temporal coherence between speech and facial motion. In this paper, we show that this strength can also become a shared vulnerability: many AV detectors appear to rely on similar cross-modal synchronization statistics as discriminative cues. Motivated by this observation, we propose a coherence-guided adversarial attack that manipulates alignment priors. Specifically, we construct a local cross-modal alignment distribution from pre-trained AV encoders and characterize it with two complementary terms: expected absolute deviation from the temporal center, which captures off-center synchronization, and alignment entropy, which captures diffuse or uncertain correspondence. We combine these terms into a unified coherence score and optimize it directly to generate bounded joint perturbations on both video and audio. Experiments on FakeAVCeleb show that the proposed method substantially degrades representative AV deepfake detectors while preserving perceptual quality. Beyond attack effectiveness, our results highlight cross-modal coherence as a potential robustness bottleneck in AV deepfake detection.