Generative artificial intelligence can now create images and videos that are almost indistinguishable from real recordings to the human eye. While this enables many useful applications, it also opens the door to misinformation, identity theft, and fraud. For this reason, the reliable detection of such deepfakes has become an important area of research.

A team from the University of Tokyo and the Max Planck Institute for Informatics has now developed a new approach that marks a significant step forward in this area. The research was conducted by researchers Kaede Shiohara and Toshihiko Yamasaki from the University of Tokyo, together with Vladislav Golyanik, Senior Researcher and head of the 4D and Quantum Vision group at the Max Planck Institute for Informatics. The team introduced their method in the paper “ExposeAnyone: Personalized Audio-to-Expression Diffusion Models Are Robust Zero-Shot Face Forgery Detectors,” which was presented at the 2026 edition of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), the world’s leading conference on computer vision.

Until now, researchers often faced a trade-off. The most accurate deepfake detectors are typically trained using supervised learning, meaning they learn from large labelled datasets containing both authentic and manipulated videos. However, such systems can become unintentionally specialized to particular forgery techniques—a phenomenon known as overfitting—and may struggle to recognize previously unseen manipulation patterns. Self-supervised approaches, in contrast, are trained exclusively on authentic videos and are therefore considered more robust against emerging deepfake technologies, but until now they have achieved lower detection accuracies. The new method is the first self-supervised method that combines robustness with the high detection rates, outperforming supervised approaches.

Rather than searching for suspicious pixel-level artifacts, the new system focuses on a person’s facial movements. It builds on the widely used FLAME model, which mathematically represents facial expressions using 53 parameters and is commonly employed in computer graphics and facial animation. The researchers pre-trained their detection model on more than 450 hours of publicly available video footage, enabling it to learn how to predict the corresponding FLAME parameters from an audio track, that is, the facial movements that would naturally be expected while speaking.

The approach can then be fine-tuned to a specific individual using only around 60 seconds of video footage, effectively turning it into a personalized deepfake investigator. During analysis, the system compares the facial movements visible in a video with the FLAME parameters predicted from the accompanying audio track. Significant discrepancies between the two are a strong indication that the video may have been manipulated.

“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,” says Vladislav Golyanik.

In experiments, the method achieved an average detection accuracy of more than 95 percent across several established benchmark datasets, surpassing the previous state of the art. Particularly challenging was an additional benchmark dataset developed by the researchers using videos generated with a frontier video-generation model, OpenAI‘s Sora 2. While previous deepfake detectors achieved results on this dataset that were barely better than a coin toss, the new method still correctly identified almost 95 percent of manipulated videos.

At the same time, the team is transparent about the limitations of their approach. The system requires extensive pre-training on powerful hardware and is currently not suitable for real-time applications. Nevertheless, they believe that their work points toward a promising direction for the next generation of deepfake detection systems.