סמינר: Graduate Seminar
Zero-shot detection of AI-generated video
Date:
October,28,2025
Start Time:
10:30 - 11:30
Location:
1061, Meyer Building
Add to:
Lecturer:
Omer Ben-Hayun
Research Areas:
| The rapid progress of generative video models has enabled the synthesis of highly realistic content, raising important concerns about misinformation and authenticity verification. However, existing supervised detection approaches often show limited generalization and scalability, as they rely on labeled data from specific generators. In this work, we present a training-free, zero-shot method for detecting AI-generated videos that does not require access to synthetic samples. Our approach leverages embedding spaces from large-scale foundation models, such as CLIP and DINOv3, and applies a whitening transformation to estimate video likelihoods in two complementary forms: spatial likelihoods, computed from per-frame embeddings, and temporal likelihoods, computed from frame-to-frame differences that capture motion dynamics. By jointly analyzing these spatial and temporal likelihoods, our method exposes systematic discrepancies in both appearance and temporal coherence between authentic and synthetic videos. Experiments across multiple benchmarks show promising trends, suggesting that this framework can serve as a robust foundation for future research in zero-shot detection of AI-generated video.
M.Sc. student under the supervision of Prof. Guy Gilboa.
|