SPACE: Source-free Proxy Anchor Concept Erasure for MLLMs

The paper introduces SPACE (Source-free Proxy Anchor Concept Erasure), a novel source-free unlearning framework for Multimodal Large Language Models (MLLMs) that addresses privacy concerns by enabling the removal of sensitive data without direct access to it. SPACE employs a two-stage process: Text-Guided Proxy Anchor Selection (TPAS) for selecting semantically aligned proxies, and Dual-Constraint Semantic Isolation (DCSI) for optimizing these anchors to erase target concepts while preserving model integrity. Experimental results demonstrate that SPACE achieves performance on par with state-of-the-art data-dependent unlearning methods across six datasets, highlighting its potential for practitioners needing effective unlearning solutions in compliance with data regulations.