Title: Roots Beneath the Cut: Uncovering the Risk of Concept Revival in Pruning-Based Unlearning for Diffusion Models (Accepted by CVPR 2026)
arxiv: https://arxiv.org/abs/2603.06640
We found pruning-based unlearning in diffusion models leak critical information that can be exploited to attack the model to recover the erased concepts. We proposed a framework to attack such unlearned models. We found that under our framework, once the locations of concept related weights are aware, the erased visual concepts can be effectively recovered. This framework has been validated effective in other diffusion unlearning methods based on concept-related weights detection and finetuning. Finally we propose a simple but effective defense strategy to conceal the locations of modified weights while maintaining the unlearning performance.
Best,
Branko
Title: Roots Beneath the Cut: Uncovering the Risk of Concept Revival in Pruning-Based Unlearning for Diffusion Models (Accepted by CVPR 2026)
arxiv: https://arxiv.org/abs/2603.06640
We found pruning-based unlearning in diffusion models leak critical information that can be exploited to attack the model to recover the erased concepts. We proposed a framework to attack such unlearned models. We found that under our framework, once the locations of concept related weights are aware, the erased visual concepts can be effectively recovered. This framework has been validated effective in other diffusion unlearning methods based on concept-related weights detection and finetuning. Finally we propose a simple but effective defense strategy to conceal the locations of modified weights while maintaining the unlearning performance.
Best,
Branko