Foundation Models are Implicit Deepfake Detectors
2026-08-10 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors found that when using pretrained self-supervised models to detect deepfakes, fake images and videos tend to have smaller feature values compared to real ones. Based on this, they treated deepfake detection as spotting anomalies and used simple feature magnitude statistics to distinguish fakes effectively. They discovered that these smaller values mostly come from changes in the meaning or content introduced by fakes rather than low-level artifacts. Additionally, larger models show stronger signals, indicating better feature representations help detect deepfakes without extra training.
Deepfake detectionPretrained modelsSelf-supervised learningFeature magnitudeAnomaly detectionSemantic shiftsGenerative fingerprintsZero-shot detectionRepresentation learningFoundation models
Authors
Stefan Smeu, Dragos-Alexandru Boldisor, Elisabeta Oneata, Dan Oneata
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.