Via lumificyber.com
The Bittensor-powered detection tool outperforms leading commercial models on key benchmarks, arriving as deepfake fraud becomes a billion-dollar problem
Deepfakes have gone from a niche internet curiosity to a nearly $900 million fraud problem in the span of a few years. BitMind Forensics, a detection system built on top of Bittensor’s decentralized AI network, is now posting benchmark scores that beat both commercial and open-source alternatives.
According to a July 2026 arXiv paper, BMF achieved an area under the curve (AUC) of 0.915 on the Deepfake-Eval-2024 image benchmark. The best commercial model in that same evaluation hit 0.90. On video detection, BMF scored 0.822, clearing the leading commercial result of 0.79.
In English: AUC is essentially a report card for how well a classifier distinguishes between real and fake content, where 1.0 is perfect and 0.5 is a coin flip.
How a crypto subnet became a deepfake hunter #
BitMind built its forensics tool on Bittensor Subnet 34, which goes by the somewhat dramatic name GAS, short for Generative Adversarial Subnet. The concept is straightforward once you strip away the jargon: miners on the network compete against each other in an adversarial loop, constantly generating and detecting synthetic media.
That cadence matters because traditional deepfake detectors are trained on a fixed dataset and deployed as static models. They work great on the fakes they’ve seen before. They tend to fall apart when confronted with outputs from newer generators. BitMind’s architecture sidesteps this problem by treating detection as a living, evolving competition rather than a one-time exam. This refresh happens roughly every four hours.
The system launched on January 15, 2025, with a mobile app designed to give users real-time deepfake detection with sub-second response times. BitMind claims the tool achieves 95% accuracy on real-world, in-the-wild content, a figure cited by co-founder and CEO Ken Jon Miyachi.
Previous deepfake detection tools averaged around 69% accuracy in real-world scenarios.
The numbers in context #
Beyond the headline benchmarks, BMF’s performance across large-scale testing is notable. The system reached a 0.936 AUC on original images from Sumsub, a well-known identity verification provider. More impressively, it posted a pooled AUC of 0.872 across a test battery exceeding 1.4 million image manipulations.
BitMind has also begun commercial integrations. CysecOnline in South Africa is among the firms that have adopted the technology, signaling early traction beyond the Bittensor ecosystem and into traditional cybersecurity channels.
Deepfake-related fraud losses totaled nearly $900 million in 2025, a figure that’s likely conservative given how many incidents go unreported or misclassified.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our