Multimodal machine unlearning · ACL 2025 Findings

MMUnlearner

Reformulating multimodal machine unlearning around visual forgetting and textual retention.

Jiahao Huo^1,3, Yibo Yan^1,2, Xu Zheng^1,2, Yuanhuiyi Lyu^1,2, Xin Zou¹, Zhihua Wei³, and Xuming Hu^1,2

¹ The Hong Kong University of Science and Technology (Guangzhou) · ² The Hong Kong University of Science and Technology · ³ Tongji University

Paper / arXiv ↗ Code ↗ Motivation & method ↓

Project overview

MMUnlearner removes target visual concepts while preserving common perception and textual knowledge.

Motivation

Multimodal unlearning requires modality-aware forgetting boundaries.

Removing a target visual concept should not inadvertently erase its textual knowledge or degrade the model's general perception and reasoning capabilities.

Machine-unlearning methods developed for unimodal classifiers do not directly specify what should be forgotten in a multimodal large language model. A target entity may be represented simultaneously through visual appearance, linguistic facts, and associations with common concepts. Updating shared parameters can therefore cause broad collateral degradation.

MMUnlearner reformulates the objective as selective visual forgetting under explicit retention constraints. It distinguishes target visual patterns from textual knowledge and common visual perception, then localizes parameter updates according to both saliency and gradient geometry.

VenueACL 2025 Findings

TaskMultimodal unlearning

Core ideaGeometry + saliency

Objectives

Selective forgetting and retention

Visual

Erase visual patterns associated with the target entity from a multimodal model.

Textual

Retain textual knowledge and language behavior that should remain available after unlearning.

Targeted

Use geometry-constrained ascent and weight saliency to localize the update.

Methodology

Saliency-guided, geometry-constrained unlearning

MMUnlearner identifies target-relevant parameters, computes competing forgetting and retention directions, and restricts optimization to a safe update subspace.

Modality-aware objectives

The training objective increases loss on target visual instances while retaining performance on textual descriptions and non-target visual concepts. This decomposition states explicitly which aspects of an entity should remain accessible.

Weight-saliency localization

A saliency map estimates the parameters most responsible for the forget set relative to retained knowledge. Optimization is concentrated on this subset to reduce unnecessary changes to the broader multimodal model.

Geometry-constrained ascent

The forgetting gradient is adjusted according to its geometric relationship with retention gradients. Conflicting components are constrained, allowing target evidence to be suppressed while limiting interference with textual knowledge and general utility.

Visual results

Framework and forgetting–retention evaluation

The method figures make the retained and forgotten regions explicit, while the experiments expose the tradeoffs.

**MMUnlearner framework**Target visual concepts are forgotten while textual knowledge and common perception are kept intact.

MMUnlearner method details — **Geometry-constrained update**Saliency and geometric constraints guide the unlearning direction and restrict harmful parameter drift.

Comparison of multimodal unlearning paradigms — **Problem formulation**Multimodal unlearning requires a different retention boundary from text-only forgetting.

MMUnlearner forgetting retention tradeoff — **Forgetting–retention tradeoff**MMUnlearner targets a stronger balance between removing the target and preserving useful capability.

MMUnlearner heatmap comparison — **Concept-level comparison**Heatmaps visualize how target and non-target concepts respond after unlearning.

MMUnlearner forgetting curve — **Forgetting curve**The evolution of target forgetting makes the effect of the constrained update visible.

Citation

BibTeX

@article{huo2025mmunlearner,
  title   = {MMUnlearner: Reformulating Multimodal Machine Unlearning in the Era of Multimodal Large Language Models},
  author  = {Huo, Jiahao and Yan, Yibo and Zheng, Xu and Lyu, Yuanhuiyi and Zou, Xin and Wei, Zhihua and Hu, Xuming},
  journal = {Findings of the Association for Computational Linguistics: ACL 2025},
  year    = {2025}
}