SAGA Tool Traces AI-Generated Videos to Their Source Systems

Researchers at the University of California, Riverside have developed SAGA (Source Attribution of Generative AI Videos), a forensic tool that identifies AI-generated videos and traces them back to the specific AI system that created them. Led by doctoral student Rohit Kundu under Professor Amit Roy-Chowdhury, in collaboration with YouTube and Google DeepMind, SAGA analyzes visual patterns or 'fingerprints' left by generative models. These patterns include spatial details within frames and temporal relationships across sequences, using a technique called Temporal Attention Signatures (T-Sigs). Tested on public datasets with 19 different AI video generators, SAGA can distinguish real from fake videos, determine whether a video was generated from text or an image, differentiate between versions of AI models, and identify the development team behind the model. The tool addresses the growing challenge of realistic AI-generated videos, shifting the critical question from whether a video is fake to what its source is.