EAI-DMCU: Evolutionary algorithm-inspired diffusion model for concept unlearning

Long Xue

doi:10.1016/j.eswa.2026.132466

EAI-DMCU: Evolutionary algorithm-inspired diffusion model for concept unlearning

2026-04-28

0

PRE

AI

DOI：10.1016/j.eswa.2026.132466

Original

Original request for help

Save

Abstract

En 中文

A text-to-image (T2I) model refers to a model that generates images conditioned on textual prompts. Diffusion models represent the most prominent class of pretrained T2I models at present. However, unwanted conceptual “knowledge” in pre-trained T2I diffusion models can result in copyright-infringing, offensive, or harmful content generation when guided by text prompts, posing ethical challenges. To this end, the concept unlearning task has been proposed, which requires that the model not only be incapable of generating undesired target concepts (i.e., forgetting effectiveness), but also preserve the generative fidelity of other non-target concepts (i.e., prior preservation). Existing methods often significantly degrade the generation fidelity of non-target concepts when forgetting target concepts, revealing substantial deficiencies in prior preservation. Inspired by evolutionary algorithms, we interpret the concept unlearning task in diffusion models from the perspective of directional selection. The concept unlearning task can be regarded as an evolutionary process conducted in a high-dimensional continuous space, subject to specific requirements of group selection. Therefore, in the concept unlearning task, the prior retention capacity of non-target concepts depends on the preservation of non-target concept features during the evolutionary process. Drawing on insights from evolutionary-algorithm analysis, we propose a new concept unlearning method, termed EAI-DMCU. We divide EAI-DMCU into an iterative evolutionary phase and a directional selection phase. During the iterative evolutionary phase, EAI-DMCU maintains invariance in cross-attention maps of target concept contexts, preserving their semantics while retaining non-target features, ultimately improving its prior preservation. In the directional selection phase, a feature-space decoupling loss is designed to unsupervisedly separate target and non-target concepts in the input prompt, thereby improving the distinction between retained and discarded individuals and enhancing the forgetting effectiveness of EAI-DMCU. By means of this procedure, EAI-DMCU guides parameter updates in the diffusion model to fulfill the concept unlearning objective. Experimental results demonstrate that EAI-DMCU exhibits superior performance across five types of concept unlearning tasks (Objects, Artstyles, Celebrities, Characters, and Nude contents), thereby validating the effectiveness of the proposed method.

Keywords:

concept unlearning

diffusion models

evolutionary algorithms

text-to-image generation

prior preservation

EAI-DMCU: Evolutionary algorithm-inspired diffusion model for concept unlearning

Abstract

Journal

Organization

Cited Papers