The authors used the demographically annotated AI-Face repository and arranged 11 deepfake face generation methods in chronological order to simulate the evolving threat landscape.
Continual learning in deepfake detectors widens demographic unfairness over time
Across 11 chronological generators, accuracy-focused continual learning methods mostly exacerbate fairness disparities, and underrepresentation alone does not explain the effect.
Independent
Kajaree Das · Oliver Böse · Subarnaduti Paul · Gabriel Storch · Martin Mundt
Research Digest··2 min read
The authors train deepfake detectors progressively on 11 real-world generators in chronological order, tracking four fairness metrics alongside accuracy.
§