"SAGA": A Tool That Detects Fake Videos and Identifies the AI Model That Generated Them

Researchers at the University of California have developed "SAGA," a tool that not only detects fake videos but also identifies the AI model that created them. It analyzes digital fingerprints and motion patterns, and was successfully tested on 19 models.

A team of researchers at the University of California, Riverside has developed a new tool called "SAGA" (Source Attribution of Generative AI Videos) to combat digital disinformation. The tool goes beyond simply detecting whether a video is fake; it can identify the specific AI model that generated it, whether based on text-to-video or image-to-video conversion. SAGA works by analyzing the unique "digital fingerprints" of each model, including motion patterns and frame sequences, using a technique called Temporal Attention Signatures. The researchers tested the tool on 19 different models, and it demonstrated the ability to verify video authenticity, determine the method of creation, identify the model used, and even pinpoint the development team. The researchers believe this tool will help investigators track disinformation campaigns and enable regulators to enforce transparency rules, despite the ongoing arms race between forgery and detection.