Saga: Source Attribution of Generative AI Videos (identifies the model used)
11 hours ago
- SAGA is the first comprehensive framework for large-scale source attribution of AI-generated videos, identifying the specific generative model used rather than just detecting real vs. fake.
- It provides multi-granular attribution across five levels: authenticity, generation task (e.g., T2V/I2V), model version, development team, and precise generator.
- The framework uses a novel video transformer architecture built on a robust vision foundation model to capture spatio-temporal artifacts.
- A data-efficient pretrain-and-attribute strategy allows SAGA to achieve state-of-the-art attribution using only 0.5% of source-labeled data per class, matching fully supervised performance.
- Temporal Attention Signatures (T-Sigs) offer a novel interpretability method that visualizes learned temporal differences, explaining why different video generators are distinguishable.
- Extensive experiments on public datasets, including cross-domain scenarios, demonstrate that SAGA sets a new benchmark for synthetic video provenance, providing interpretable insights for forensic and regulatory applications.