ML Security PapersFusion-SSAT: Unleashing the Potential of Self-supervised Auxiliary Task by Feature Fusion for Generalized Deepfake Detection
Shukesh Reddy 1, Srijan Das 2, Abhijit Das 1
Published on arXiv
2601.00789
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Fusing self-supervised auxiliary task features with primary encoder features yields better cross-dataset generalization than current state-of-the-art deepfake detectors across six benchmark datasets
Fusion-SSAT
Novel technique introduced
In this work, we attempted to unleash the potential of self-supervised learning as an auxiliary task that can optimise the primary task of generalised deepfake detection. To explore this, we examined different combinations of the training schemes for these tasks that can be most effective. Our findings reveal that fusing the feature representation from self-supervised auxiliary tasks is a powerful feature representation for the problem at hand. Such a representation can leverage the ultimate potential and bring in a unique representation of both the self-supervised and primary tasks, achieving better performance for the primary task. We experimented on a large set of datasets, which includes DF40, FaceForensics++, Celeb-DF, DFD, FaceShifter, UADFV, and our results showed better generalizability on cross-dataset evaluation when compared with current state-of-the-art detectors.
Key Contributions
- Fusion-SSAT architecture that combines a self-supervised auxiliary reconstruction task with a primary deepfake detection task via feature fusion, blending local texture and global semantic representations
- Empirical study of different training scheme combinations for auxiliary and primary tasks, finding feature-level fusion to be most effective
- Demonstrated improved cross-dataset generalization across seven benchmarks (DF40, FaceForensics++, Celeb-DF, DFD, FaceShifter, UADFV) compared to state-of-the-art detectors
🛡️ Threat Analysis
Proposes a novel detection architecture for identifying AI-generated/manipulated facial content (deepfakes) — directly falls under output integrity and AI-generated content detection. The primary contribution is a new detection method (Fusion-SSAT), not merely an application of existing methods.