Thread Details: Continual Training Across 11 Deepfake Generators, 4 Fairness Metrics
mundt_martin · x · 2026-10-25
Author mundtmartin's thread post #2 detailing the experimental design of the preprint "Time to Fake It": 11 deepfake datasets ordered by release date (FaceForensics++ 2019 to Stable Diffusion-based fakes 2023), detectors continually trained in sequence, six continual learning methods compared, and 4 fairness metrics measured after every step.
Same paper as the main announcement, with the consistent finding that continual learning preserves accuracy at the cost of fairness across age, gender, and skin tone.
More from Research
- MEMI Reverse-Engineers Robot Memory, Hits 97% on RoboMME (99.4% with Controller) — du_yilun · 2026-10-25
- Microsoft-led BMJ paper defines "medical superintelligence": isolated benchmark wins aren't enough — EricTopol · 2026-10-25
- AI Writing Is Getting Harder to Read — Research Says It's Trainable — lmoroney · 2026-10-25
- LLMail-Inject: 208K real prompt injection attacks against LLM email assistants — niloofar_mire · 2026-10-25
- Paper2Agent turns scientific papers into AI research assistants, published in Nature — james_y_zou · 2026-10-25
- 14.8M-prompt EMNLP study: different demographic cues flip LLM bias conclusions — lucy3_li · 2026-10-25