• 2 min read
New deepfake detector tops 95% by reading facial motion
Researchers from the University of Tokyo and Max Planck say their self-supervised system catches deepfake videos with over 95% accuracy.
Image: TechXplore
A new deepfake detection method from the University of Tokyo and the Max Planck Institute for Informatics reports more than 95% average accuracy by focusing on whether a person’s facial movements match the audio, instead of hunting for visual glitches.
The work, by Kaede Shiohara, Toshihiko Yamasaki, and Vladislav Golyanik, was presented at CVPR 2026 in a paper titled “ExposeAnyone: Personalized Audio-to-Expression Diffusion Models Are Robust Zero-Shot Face Forgery Detectors.” The team says the approach handled manipulations that caused many existing detectors to fail.
How ExposeAnyone works
Most high-performing detectors rely on supervised learning with large labeled sets of real and fake videos. That can make them effective on known forgery methods, but also prone to overfitting. Self-supervised systems, trained only on authentic footage, are generally seen as more robust to new deepfake techniques, though they have usually lagged on accuracy.
Recommended reading
Gemini CLI helped run an 8-device botnet
This system tries to close that gap. It uses the FLAME face model, which represents expressions with 53 parameters. The researchers pre-trained their model on more than 450 hours of public video to predict expected FLAME parameters from speech audio. They then fine-tuned it for a specific person using only around 60 seconds of reference video.
When analyzing a suspicious clip, the detector compares the facial movements visible in the video against the movements the audio would naturally imply. Large mismatches suggest manipulation.
“The combination of self-supervised learning and FLAME-based facial analysis makes our approach particularly robust against new deepfake generation methods as well as distortions such as image compression or noise.”
Sora 2 benchmark results and limits
Across established benchmark datasets, the team reports more than 95 percent average detection accuracy, outperforming prior methods. On an additional dataset the researchers built using videos generated with OpenAI’s Sora 2, they say earlier detectors performed only barely better than a coin toss, while the new method still correctly flagged almost 95 percent of manipulated videos.
The tradeoff is compute. The system needs extensive pre-training on powerful hardware and, for now, is not suitable for real-time use.
The paper is available on arXiv with DOI 10.48550/arxiv.2601.02359.
Security Editor
Sophia unpacks the invisible wars happening on our networks. Covering cybersecurity, privacy legislation, and cryptography, she exposes how our data is weaponized and defended. Before joining for(geeks), she spent years as a penetration tester. She's the reason the rest of the team uses physical security keys.
via TechXplore
