ERASED OR DORMANT? RETHINKING CONCEPT ERASURE THROUGH REVERSIBILITY

ICLR 2026 Conference Submission9995 Authors

Keywords: concept erasure; diffusion model

Abstract:

To what extent do concept erasure techniques in diffusion models truly remove, rather than merely suppress, targeted concepts?
In this paper, we explore this question by introducing a diagnostic framework that leverages lightweight parameter adaptation to probe the robustness and reversibility of leading erasure methods.
Central to our approach are two minimal yet general probes: (i) a Gradient-Guided Probe, which restores suppressed behavior by reversing gradient signals, and (ii) an Instance-Personalization Probe, which reinstates concepts through few-shot supervision.
Across six erasure algorithms, multiple concept types, and diverse diffusion backbones, we consistently find that erased concepts can be recovered with high fidelity after only minimal adaptation.
Our theoretical analysis reinforces these results, showing that reversed weight remain bounded to the original parameters, leaving much of the targeted representation intact.
Together, these findings demonstrate that existing methods do not eliminate concepts but merely push them below the surface, where they can be readily revived. As such, our work calls for a rethinking of concept erasure: moving beyond superficial suppression toward approaches that dismantle latent structures at their core, alongside more rigorous standards for evaluating safety in generative models.

Primary Area: generative models

Submission Number: 9995