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

Deepfake detection has historically suffered from a structural failure: models are trained on a fixed set of forgeries and then deployed into an environment where generative techniques evolve weekly. This leads to rapid performance decay, with state-of-the-art detectors often losing 45-50% of their accuracy when tested on real-world content circulating on social media. The authors argue that the field's reliance on static benchmarks is the primary cause of this "generalization gap."

To address this, the authors introduce BitMind Forensics (BMF), a system built on the Bittensor SN34 network. This is an open, adversarial competition where participants act as either "miners" (creating new deepfakes) or "detectors" (building models to catch them). This mechanism creates a continuous feedback loop: as new generative models emerge, they are added to the training distribution, forcing the detector population to adapt or lose economic incentives. The production model is a snapshot of this evolving process, utilizing a heterogeneous ensemble of vision backbones (ConvNeXt, EVA, CLIP, and DINOv3) to ensure robustness across different architectural families.

BMF was evaluated across nineteen public datasets, including canonical face-swap suites and modern AI-generated media benchmarks. The results demonstrate that BMF consistently outperforms static baselines. On the Deepfake-Eval-2024 benchmark, BMF matches the best commercial detectors on images and exceeds them on video. Furthermore, the authors show that successive snapshots of the system improve over time, proving that the continuous training process successfully tracks the moving generative frontier. The system also maintains high performance under common manipulations like JPEG compression and downscaling, which typically defeat standard models.