New Tool Traces AI Videos Back to Their Source

Researchers at UC Riverside have developed SAGA, a tool designed to trace the origin of AI-generated videos and identify the specific generative model used to create them. The tool goes beyond simple detection, reasoning about why a video is fake and identifying the source. The team hopes to foster collaboration among model developers to proactively address the growing problem of AI-generated disinformation and deepfakes by alerting them to models being used to create harmful content. The tool utilizes temporal signatures – unique patterns in how frames evolve over time – to distinguish between different generative models.

SAGA works by analyzing temporal signatures, which are unique patterns in how frames evolve over time. Different generative models produce distinct temporal artifacts, meaning that even if two models are given the same prompt, their outputs will exhibit subtle differences in how frames change over time. This allows SAGA to identify which model was used to generate a specific video.

The researchers acknowledge that AI can still struggle to reliably identify fake videos, and the challenges will only increase as SAGA is deployed in the real world. To mitigate this, they built the tool on top of a foundation model, which prevents it from being thrown off by common variations in data. They also demonstrated the effectiveness of their video transformer architecture, specifically tailored for video attribution.

Collaboration is a key component of the researchers’ vision. They want to create a more collaborative environment where model developers are alerted when their technology is being used to generate harmful content, such as deepfakes of celebrities or politicians used for disinformation campaigns, or business email compromise attacks impersonating executives. By sharing information about frequently flagged models, developers can implement safeguards and restrictions to prevent the generation of unsafe content in the first place.