Training-Free AI Video Attribution via Retrieval and Generative Fingerprints
A new paper on arXiv introduces an innovative method for pinpointing the sources of AI-generated videos without needing training. Named 'Retrieval-Driven Training-Free AI-Generated Video Attribution' (arXiv:2607.28955), it addresses the limitations of existing techniques that primarily focus on images and rely on image generation models, which struggle with larger video datasets. This new approach frames video attribution as a retrieval problem and incorporates a modified orthogonal color transformation, multi-scale quantized residual generation, and temporal-semantic aggregation. By enhancing the ability to trace the origins of realistic AI-generated videos, this method aims to support forensic investigations and legal governance, especially in the face of cybersecurity challenges. The paper was released as a cross-type submission and is available at the given URL.
Key facts
- Paper title: Retrieval-Driven Training-Free AI-Generated Video Attribution
- arXiv ID: 2607.28955
- Announce type: cross
- Method is training-free
- Formulates attribution as instance retrieval
- Pipeline includes adapted orthogonal color transformation, multi-scale quantized residual generation, and temporal-semantic aggregation
- Addresses limitations of image-based attribution methods
- Aims to improve forensic investigation and legal regulation
- Motivated by threats to cybersecurity and social governance
- Available at https://arxiv.org/abs/2607.28955
Entities
Institutions
- arXiv