PRMU: A Corpus-Free Benchmark for Person-Centric Knowledge Unlearning in Multimodal Large Language Models

2026-08-11

AI Summary

This work addresses the challenge of reliably unlearning person-specific knowledge in multimodal large language models when the original training data is unavailable. To this end, it introduces PRMU, the first person-centric, training-data-free benchmark for evaluating multimodal unlearning, and proposes SGPE, a lightweight parameter-editing method. SGPE integrates similarity-gated projection editing, adversarial probing, and fine-grained locality analysis to achieve effective unlearning while preserving knowledge locality and general model capabilities. Experimental results demonstrate that existing approaches often suffer from degraded locality under aggressive unlearning and are vulnerable to reactivation via multimodal cues, whereas SGPE achieves a superior trade-off among unlearning efficacy, knowledge retention, and multimodal robustness.

Abstract

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in storing and recalling rich person-related knowledge, raising increasing concerns about reliable knowledge removal. However, existing machine unlearning approaches for MLLMs typically assume access to original forget and retain corpora, which are often unavailable in realistic deletion scenarios. To address this limitation, we introduce PRMU, a benchmark for evaluating corpus-free multimodal unlearning under realistic person-centric deletion requests. PRMU focuses on naturally acquired person-related knowledge and evaluates whether models can remove target knowledge while preserving related knowledge through diverse textual and visual probes, including adversarial evaluation and fine-grained locality analysis. To facilitate research in this setting, we further introduce Similarity-Gated Projection Editing (SGPE), a lightweight corpus-free unlearning baseline with knowledge displacement, protected parameter-space editing, and locality-aware multimodal control. Extensive experiments on representative MLLMs reveal that existing unlearning methods often suffer from unfavorable forgetting-locality trade-offs, with significant locality degradation under aggressive forgetting settings, and remain vulnerable to multimodal knowledge reactivation. Meanwhile, SGPE provides a competitive trade-off between target forgetting, locality preservation, and general multimodal utility. We hope PRMU can facilitate future research toward realistic and scalable multimodal machine unlearning. Code and dataset will be released at https://github.com/2231122/PRMU.

Problem

• multimodal unlearning
• person-centric deletion
• corpus-free
• knowledge removal
• large language models

Innovation

• corpus-free unlearning
• multimodal large language models
• person-centric knowledge removal
• Similarity-Gated Projection Editing
• locality preservation