ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition
For practitioners needing fine-grained control over what VLMs forget, this work provides a method to remove specific concepts without affecting unrelated image content, addressing a key limitation of instance-level unlearning.
The paper tackles the problem of imprecise knowledge removal in Vision-Language Models (VLMs) when unlearning at the instance level. It proposes ICED, a framework that decomposes visual representations into interpretable concepts, enabling selective suppression of target concepts while preserving non-target semantics, achieving more comprehensive forgetting and better model utility than existing methods.
Machine unlearning in Vision-Language Models (VLMs) is typically performed at the image or instance level, making it difficult to precisely remove target knowledge without affecting unrelated semantics. This issue is especially pronounced since a single image often contains multiple entangled concepts, including both target concepts to be forgotten and contextual information that should be preserved. In this paper, we propose an interpretable concept-level unlearning framework for VLMs, which constructs a compact task-specific concept vocabulary from the forgetting set using a multimodal large language model. In addition to modality alignment, visual representations are decomposed into sparse, nonnegative combinations of semantic concepts, providing an explicit interface for fine-grained knowledge manipulation. Based on this decomposition, our method formulates unlearning as concept-level optimization, where target concepts are selectively suppressed while intra-instance non-target semantics and global cross-modal knowledge are preserved. Extensive experiments across both in-domain and out-of-domain forgetting settings demonstrate that our method enables more comprehensive target forgetting, better preserves non-target knowledge within the same image, and maintains competitive model utility compared with existing VLM unlearning methods.