Seminar: Signal Processing and Systems
Manifold Induced Biases for Detection of Generated Images and their Memorization
Date:
March,19,2025
Start Time:
14:30 - 15:30
Location:
1061, Meyer Building
Add to:
Lecturer:
Jonathan Brokman
Research Areas:
| In this seminar, we present our ICLR 2025 paper, “Manifold Induced Biases for Zero shot and Few-shot Detection of Generated Images” where we apply total-variation (TV) analysis to AI-generated content detection and memorization in a novel zero- shot framework.
Distinguishing real from AI-generated images remains a challenge. While supervised methods dominate the field, zero-shot approaches offer a scalable alternative, eliminating the reliance on datasets which quickly become out-dated as generative models evolve. We leverage biases in the implicit probability manifold of pre-trained diffusion models, using score-function analysis to approximate the TV sub-gradient, gradient magnitude, and manifold bias—leading to effective zero-shot detection criteria. We will also briefly discuss our papers published in TOG 2024, ICLR 2024 and SSVM 2025. Ph.D. student under the supervision of Prof. Guy Gilboa.
|