AI-generated videos are becoming increasingly difficult to distinguish from authentic footage, creating new challenges for digital forensics. Researchers led by a team at the University of California, Riverside, have developed SAGA, a framework that not only detects synthetic videos but also identifies the AI system that produced them, tells Tech Xplore.
SAGA, short for Source Attribution of Generative AI Videos, looks for subtle visual patterns unintentionally left behind by video generators. These artifacts act like fingerprints, allowing researchers to distinguish content created by different generative models. The project was led by UC Riverside doctoral student Rohit Kundu under professor Amit Roy-Chowdhury, in collaboration with researchers from YouTube and Google DeepMind.
Unlike image analysis, video forensics can draw on both spatial and temporal information. SAGA examines details within individual frames while also tracking how visual information changes across a video sequence. Its key technique, called Temporal Attention Signatures, or T-Sigs, captures patterns associated with individual generators. By averaging signatures from multiple videos created by the same system, researchers can build a characteristic profile for that generator.
The team evaluated SAGA using public datasets containing videos from 19 AI video generators. The collection included text-to-video models, which generate footage from written prompts, and image-to-video models, which animate still images.
Testing showed that SAGA can perform several levels of attribution. It can determine whether footage is real or synthetic, identify whether an AI video originated from text or an image, differentiate versions of underlying models, and even identify the development team responsible for a model.
Source attribution could give investigators, regulators, and technology companies a stronger tool for tracking synthetic media. As AI-generated video becomes more convincing and widespread, identifying its origin could help investigate misinformation, fraud, and other deceptive uses.