Jing Wu, Trung Le, Mingyi Zhou, Dinh Phung, Jianfei Cai, Mehrtash Tafazzoli Harandi · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3836000
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Machine unlearning methods have been proposed to delete undesired knowledge from Text-to-Image diffusion models (DMs), yet, unlearned models remain vulnerable to adversarial prompts that regenerate erased content. While existing noise-prediction optimization attacks have shown promising results, they generally rely on the intuitive premise that similar noise predictions between the original and unlearned models will naturally yield similar generated images. In this paper, we observe that this noise-level alignment may not fully guarantee image-level consistency, primarily due to cumulative deviations between noise prediction and image reconstruction during the iterative denoising process. Building on this diagnosis, we introduce a latent consistency loss and integrate it into existing optimization-based attack frameworks to bridge this gap. We further examine the vulnerability of unlearned DMs under more strict settings, and propose a gradient-free attacker (ZOOAtk) via zeroth-order stochastic variance reduction to enable effective prompt optimization under limited access. Experimental results demonstrate that latent consistency strengthens attacks in different settings, while ZOOAtk remains effective under restricted access, indicating that current unlearned DMs remain vulnerable even when internal model information is unavailable.
No comments yet — start the discussion below.