Adversarial Attack Framework Against Vision-Language Model Unlearning

摘要

Large Vision--Language Models (VLMs) unlearning tends to eliminate the influence of ``to-be-forgotten'' content in the training corpora, algorithmically by suppressing the likelihood of faithfully generating responses on forget-target inputs. The injection of adversarial inputs can manipulate the unlearned VLM's generation towards the attacker’s will, forcing the reproduction of the supposedly forgotten content and undermining the reliability of expected forgetting behavior. However, most attacks assume access to the unlearned VLM’s architecture or parameters, or to output logits via queries. In this paper, we propose SISA, a novel attack framework for crafting adversarial inputs to manipulate generation towards the forgotten target, which only requires access to a surrogate, pre-trained VLM. SISA advances prior attacks by exploiting the persistence of visual sink tokens after unlearning as a stable structural anchor for semantic alignment. SISA induces a sink regime on a candidate visual token to build the structural anchor that influences generation, and then semantically aligns model output to the target while conditioning on attention through the induced sink token, reinforcing the anchor for desired elicitation. With sink persistence and sink-conditioned semantic anchoring, SISA crafts transferable adversarial inputs. Evaluation on diverse unlearned VLM settings confirms the effectiveness of SISA, increasing the outputs’ semantic agreement with forgotten targets by up to 6.9X relative to clean inputs.