Dynamic Adversarial Training Matches Commercial Deepfake Detectors

markjeffrey · x · 2026-07-28

Current Deepfake detectors excel on academic benchmarks but experience a 45-50% AUC drop in real-world scenarios. To solve the issue of static models falling behind moving generative frontiers, BitMind introduced BitMind Forensics (BMF).

BMF is continuously refreshed through Bittensor SN34, an open adversarial competition. Evaluated across 19 public datasets, BMF achieved 0.915 AUC on images and 0.822 AUC on videos in the Deepfake-Eval-2024 benchmark, outperforming the best commercial video detectors while maintaining strong robustness against perturbations.

Related event: BitMind's Dynamic Adversarial Training Matches Commercial Deepfake Detectors(2 posts)→

Original post →

More from Safety

Safety channel →