BitMind Forensics beats deepfakes with continuous training, scoring 0.991 AUC

BitMind Forensics (BMF) is a deepfake detection system that never stops learning. Built on Bittensor SN34—a decentralized adversarial competition—it constantly refreshes its training data as new generative models emerge. This solves a critical flaw in static detectors: they achieve near-perfect scores on academic benchmarks but suffer 45-50% AUC drops on real-world content because they were trained once against a fixed generative frontier.

BMF was evaluated across nineteen public datasets spanning classic face-swaps (FaceForensics++, Celeb-DF, DFDC) and recent in-the-wild benchmarks (Sumsub, Deepfake-Eval-2024, WildRF, RAID, GenVideo-100K). On Sumsub's full four-condition manipulation battery (1.4M images), BMF achieves a pooled AUC of 0.872, with 0.855 under JPEG compression and 0.799 under downscaling—far above open-source alternatives. On Deepfake-Eval-2024, it matches the best commercial detector on images (0.915 vs 0.90) and beats it on video (0.822 vs 0.79), while open-source detectors lag at 0.56 and 0.63.

On the AI-image front, BMF reaches 0.991 AUC over 21 generators and 0.918 on GenVidBench. It also exceeds the FF++-trained frontier on contamination-audited DFDC (0.947 vs 0.843) and Celeb-DF v2 (0.9985 vs 0.956). A temporal study shows successive dated exports improve on held-out media from unseen generators (image AUC from 0.842 to 0.902; video from 0.864 to 0.936), proving continuous learning works. All evaluation code and the production API snapshot are public for independent verification.