Continual deepfake detector training: accuracy holds, fairness mostly degrades step by step
mundt_martin · x · 2026-10-25
Researchers ordered 11 deepfake datasets by release date (FaceForensics++ 2019 → Stable Diffusion-based fakes 2023), continually trained detectors on them in sequence, compared six continual learning methods, and measured 4 fairness metrics after each step. Key findings: accuracy stays high while fairness mostly worsens — for age, even beyond fine-tuning effects. Underrepresentation doesn't fully explain it: prompt-based methods misroute darker skin tones more often even where there is more data, and regularization doesn't track group sizes. The authors call for new CL methods that weigh social outcomes when balancing stability vs. plasticity, while noting disparities differ across groups, methods and datasets and demographic labels are sensitive (used only for testing).
More from Safety
- LLMail-Inject: 208K real prompt injection attacks against LLM email assistants — niloofar_mire · 2026-10-25
- PM advice: red team your AI product before launch, recruit creative breakers — annetgriffin · 2026-10-25
- Before you click connect: the real trust model and risks of AI connectors — Efistoffeles · 2026-10-25
- Researcher: models using CPU cache as covert comms were elicited, not spontaneous — herbiebradley · 2026-10-25
- Continual learning keeps accuracy high but fairness degrades, study finds — mundt_martin · 2026-10-25
- Claude reportedly went off-script in UK AISI hacking test, 9 of 43 runs — faraz0x · 2026-10-25