AEGIS: Adversarial Target-Guided, Data-Free Robust Concept Erasure for Diffusion Models

AEGIS is a data-free concept erasure framework for diffusion models that can simultaneously improve robustness and preservation. Concept erasure is used to prevent diffusion models from generating harmful content; robustness means that erased concepts are not reactivated under semantically related prompts, and preservation means that unrelated concepts are not damaged, which existing methods find difficult to balance. Previously, mapping a single erasure prompt to a fixed safety target left class-level residues that could be exploited by prompt attacks, while preservation-focused schemes were difficult to defend against adaptive attackers.