Saga: Source Attribution of Generative AI Videos (identifies the model used) (arxiv.org)

Researchers have introduced SAGA (Source Attribution of Generative AI videos), a groundbreaking framework designed to identify the specific generative models used to create hyper-realistic synthetic videos. This development is significant for the AI/ML community as it addresses the growing challenge of distinguishing between real and fabricated video content, especially in an era where misuse of generative AI poses serious risks. SAGA goes beyond traditional binary detection methods, offering multi-granular source attribution across five distinct levels, including authenticity and the specific model version, thereby enhancing forensic analysis capabilities. SAGA employs a novel video transformer architecture that captures essential spatio-temporal artifacts from video data and introduces a data-efficient pretrain-and-attribute strategy, achieving state-of-the-art performance with only 0.5% of labeled data per class. It also features a new interpretability method, Temporal Attention Signatures (T-Sigs), which provides visual insights into the temporal differences that differentiate various video generators. Through extensive testing across diverse datasets, SAGA sets a new benchmark for synthetic video provenance, making it a vital tool for forensic applications and regulatory compliance in the rapidly evolving field of generative AI.