EE ZOOM Seminar: Low-Resource Reconstruction of Template-Memorized Images — Attack against Diffusion Models
https://tau-ac-il.zoom.us/j/85063969240?pwd=t2xMC9bTXyQ2mtCTfTkywVCGJ8SvFc.1
Meeting ID: 850 6396 9240
Passcode: 529492
Electrical Engineering Systems Seminar
Speaker: Sol Yarkoni
M.Sc. student under the supervision of Prof. Roi Livni
Wednesday, 18th June 2025, at 15:00
Low-Resource Reconstruction of Template-Memorized Images — Attack against Diffusion Models
Abstract
Diffusion models trained on large-scale datasets such as LAION have become foundational to modern generative AI. However, their reliance on uncurated web data introduces serious privacy risks, particularly through the unintended extraction of training images. This seminar presents the findings of the research paper “Low Resource Reconstruction Attacks Through Benign Prompts”, recently submitted to NeurIPS and an ICML workshop.
The talk introduces a low-resource image reconstruction attack capable of extracting template-memorized content from diffusion models using only black-box access and natural English prompts such as “Abstract Art T-Shirt.” The attack successfully reconstructs training images, including some containing real, identifiable individuals.
To contextualize the attack, the talk will also review the phenomenon of template memorization and examine how data scraped from e-commerce websites contributes to the image–text coupling hypothesized to underlie this behavior.
Sol Yarkoni is an M.Sc. student, experienced data scientist in computer vision, and generative artist.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום בצ'ט של שם מלא + מספר ת.ז.
