To address the problem that artificial intelligence can generate videos so realistic that distinguishing them from authentic footage is becoming increasingly difficult, a computer science team led by researchers at UC Riverside has developed a tool that moves beyond simply identifying whether a video is fake. It also determines which AI system created it.
The new framework identifies visual patterns within fake video frames that were unintentionally introduced by AI video generators, providing a forensic path to identifying the systems that created the videos.
“The patterns are like fingerprints that the generative model leaves behind, and our goal here was to find out if the signatures are distinct amongst different generators. It turns out that, yes, there are distinct fingerprints there,” said UCR doctoral student Rohit Kundu, who led the research under the guidance of Professor Amit Roy-Chowdhury in collaboration with researchers at YouTube and Google DeepMind.
The new framework is called SAGA — short for Source Attribution of Generative AI Videos. It represents one of the first large-scale efforts to trace AI-generated videos back to their origins.
The innovation comes as generative AI systems are rapidly improving and synthetic videos are increasingly being used in entertainment, advertising, education, social media, and politics. At the same time, concerns are growing about their misuse in misinformation campaigns, fraud, and other deceptive activities.
“Knowing that a video is fake is often not enough,” the researchers wrote. “The critical need has shifted from whether it’s fake to what is its source?”
The ability to identify the source of AI-generated content could help digital forensic investigators track misinformation campaigns, assist regulators in enforcing transparency requirements, and aid technology companies in understanding how synthetic media spreads online.
Unlike still images, videos contain motion and temporal information with visual elements changing from one frame to the next over time. Different AI video generators produce subtle patterns in those changes, creating distinctive artifacts.
To detect those artifacts, the team analyzed both spatial details within individual frames and temporal relationships across entire video sequences. Rather than examining a single frame, the system studies how visual information evolves from moment to moment, the researchers explained.
A key innovation is a technique called Temporal Attention Signatures, or T-Sigs. These signatures visualize the unique patterns associated with different video generators. By averaging patterns across many videos produced by the same AI system, SAGA generates a characteristic profile that helps distinguish one generator from another.
The researchers tested SAGA using public datasets containing videos generated by 19 different AI video generators. These included both text-to-video systems, which create videos from written prompts, and image-to-video systems, which animate still images.
The system can determine whether a video is real or synthetic, identify whether it was created from text or images, distinguish among different versions of underlying AI models, and even identify the development team behind a model.
