Resemble AI expands multimodal deepfake detection approach for new accuracy highs

The field of deepfake detection has a problem: Detectors are much less effective on deepfakes created by generators newer than they are, as well as unfamiliar models. Resemble AI says by rethinking what is being detected, it has moved beyond this performance limitation with its new deepfake detection model.

Researchers with the Vector Institute have shown “performance drops sharply” if the deepfake detector has not been trained on content from the generator that created a given deepfake, according to a Resemble AI post explaining its new DETECT-World software.

Instead of training a deepfake detection model to recognize artifacts found in the creations of known generators, the company says it built its model to consider  “does this content violate my model of how physical reality works?”

The company claims 99.47 percent audio deepfake detection accuracy, 98.2 percent detection accuracy for video and 95.8 percent for images.

DETECT-World is the third generation deepfake detection model developed by Resemble AI, following DETECT-2B and DETECT-3B Omni, the latter of which added image and video capabilities. It is now operating within Resemble Detect, the company’s real-time deepfake detection API.

The change is in implementing the World-Vision Hybrid Encoder to look for inconsistency with the physical world, and return a manipulation probability score, per-frame scores and a spatial heatmap.

An example of its effectiveness, according to the post, is seen in DETECT-World’s 95 percent accuracy catching Haotian-style real-time face-swap attacks. These attacks were misclassified by leading detectors at rates near 100 percent in an investigation published by 404 Media in May. The new model had not been exposed to this type of attack before, Resemble says.

Mountain View, California-based Resemble AI was sitting second on Hugging Face’s Speech Deepfake Arena Leaderboard when the benchmark was paused for its next stage.

Article Topics

accuracy  |  deepfake detection  |  deepfakes  |  research and development  |  Resemble AI