Continuously Evolving Deepfake Detection: An Architecture and Public-Benchmark Evaluation of a Dynamic Detection System

arXiv cs.AI · 18 hours ago

Researchers developed BitMind Forensics, a deepfake detection system trained through continuous adversarial competition via Bittensor SN34, to address the gap where static detectors drop 45-50% in performance on real-world content despite high academic benchmark scores. The system achieved 0.936 AUC on Sumsub's original images, 0.822 AUC on Deepfake-Eval-2024 video (exceeding commercial detectors at 0.79), and 0.991 AUC on AI-generated images across nineteen public datasets. By continuously refreshing training data against evolving generative methods, successive model exports improved performance from 0.842 to 0.902 AUC on unseen image generators and 0.864 to 0.936 on video generators.