🤖 AI Summary
This work investigates the intrinsic mechanisms by which foundation models detect deepfakes, addressing why pretrained representations effectively distinguish authentic from synthetic media. The study reveals that forged samples consistently elicit lower-magnitude feature responses across diverse foundation models and systematically demonstrates— for the first time—that this amplitude discrepancy serves as a key signal for authenticity verification, rooted in semantic shift. Building on this insight, the authors reformulate deepfake detection as an anomaly detection task, showing that simple statistics of feature magnitudes alone enable efficient zero-shot detection. The proposed approach achieves performance on par with complex specialized models across both image and video modalities, with detection capability scaling favorably with model size, thereby confirming that large-scale foundation models inherently possess strong zero-shot potential for forgery identification.
📝 Abstract
Pretrained self-supervised representations have emerged as a core component of current deepfake detection methods, yet it remains unclear which of their properties make real and fake media distinguishable. In this work, we uncover a surprisingly consistent phenomenon: across multiple pretrained models, datasets, and both image and video domains, fake samples systematically produce lower-magnitude representations than their real counterparts. Motivated by this finding, we formulate deepfake detection as an anomaly detection problem and show that simple statistics of feature magnitude achieve competitive performance with far more sophisticated deepfake detection methods. We further investigate the origin of this effect and demonstrate that reduced feature magnitude is primarily associated with semantic shifts introduced by fake content, while low-level generative fingerprints play a comparatively smaller role. Finally, we show that this discriminative signal strengthens as the size of the underlying foundation model grows, suggesting that advances in representation learning naturally translate into stronger zero-shot deepfake detectors.