Survey Explores Adversarial Attacks and Defenses in Diffusion Models

A new preprint published on arXiv presents a unified survey on the security of diffusion models across multiple visual domains. The study analyzes adversarial attacks and defensive techniques targeting generative AI systems.

The review covers three primary visual modalities: images, videos, and 3D data. The authors introduce a task-centric taxonomy, categorizing existing literature by modality, generative task, and chronological progression of attack and defense mechanisms.

Furthermore, the paper consolidates the evaluation benchmarks, datasets, and performance metrics used in current research. The authors conclude by outlining open challenges and proposing key directions for future exploration in generative model security.