Continual Learning Keeps Deepfake Detection Accurate but Worsens Fairness, Study Finds

The OWL-ML team at the University of Bremen released the preprint "Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection," with author Martin Mundt and colleagues explaining the study design and findings in a tweet thread.

Confirmed

Experimental design: 11 deepfake datasets were ordered by release date—from FaceForensics++ in 2019 to fake-face data generated with Stable Diffusion in 2023—so the detector could be continually trained in sequence, comparing the performance of six methods (not fully listed in the post).
Core finding: after continual learning, the detector's accuracy remained high, but most fairness metrics deteriorated; on the age dimension it performed even worse than the fine-tuning baseline.
The authors note that inter-group disparities vary by method and dataset, and that demographic labels are themselves sensitive information, which the team used only for testing.
The authors call for new continual learning methods to account for societal outcomes when balancing stability and plasticity, while admitting that it is "easier said than done."

Why it matters

In real-world deployment, deepfake detection models must continually adapt to new generators. This study shows that simply chasing accuracy widens the performance gap across demographic groups—a societal cost.
The results suggest future continual learning research should incorporate fairness into evaluation and method design, rather than looking only at overall accuracy.

2026-10-25 ~ 2026-10-25 · 5 related posts

Primary sources

Continual deepfake detector training: accuracy holds, fairness mostly degrades step by step — mundt_martin · 2026-10-25 23:26 (UTC+08:00)
Continual learning keeps accuracy high but fairness degrades, study finds — mundt_martin · 2026-10-25 23:26 (UTC+08:00)
Continual Learning Keeps Deepfake Detectors Accurate but Worsens Fairness, Study Finds — mundt_martin · 2026-10-25 23:26 (UTC+08:00)

[source] Continual Learning Keeps Deepfake Detectors Accurate but Worsens Fairness, Study Finds — mundt_martin · 2026-10-25
Thread Details: Continual Training Across 11 Deepfake Generators, 4 Fairness Metrics — mundt_martin · 2026-10-25
[source] Continual deepfake detector training: accuracy holds, fairness mostly degrades step by step — mundt_martin · 2026-10-25
[source] Continual learning keeps accuracy high but fairness degrades, study finds — mundt_martin · 2026-10-25
Team Credits and Caveats: Fairness Labels Are Sensitive, Disparities Vary Widely — mundt_martin · 2026-10-25