Stable Multimodal Graph Unlearning via Feature-Dimension Aware Quantile Selection

2026-05-04

AI Summary

Existing graph unlearning methods apply a uniform editing strategy across all GNN layers in multimodal graphs, often over-modifying high-dimensional input projection layers and thereby significantly degrading model utility while struggling to balance privacy and stability. To address this, this work proposes the FDQ framework, which introduces, for the first time, a feature-dimension-aware mechanism to adaptively identify sensitive layers. By integrating diagonal sensitivity analysis with quantile-based threshold selection, FDQ applies more conservative editing strategies to high-dimensional layers during suppression set construction. Notably, this approach enables lightweight, plug-and-play, layer-wise differentiated parameter editing without altering the underlying importance evaluation mechanism. Experiments demonstrate that FDQ effectively mitigates membership inference attacks on the Ele-Fashion and Goodreads-NC datasets while substantially outperforming baseline methods in preserving model utility.

Abstract

Graph unlearning remains a critical technique for supporting privacy-preserving and sustainable multimodal graph learning. However, we observe that existing unlearning strategies tend to apply uniform parameter selection and editing across all graph neural network (GNN) layers, which is especially harmful for multimodal graphs where high-dimensional input projections encode dominant cross-modal knowledge. As a result, over-editing these sensitive layers often leads to catastrophic utility degradation after forgetting, undermining both stable learning and effective privacy protection. To address this gap, we propose FDQ, a Feature-Dimension Aware Quantile framework for multimodal graph unlearning. FDQ adaptively identifies high-dimensional input projection layers and applies more conservative, FDQ-guided quantile thresholds when constructing suppression sets, while keeping the underlying importance estimation mechanism unchanged. FDQ is seamlessly integrated with diagonal sensitivity-based parameter importance analysis to enable efficient node and edge unlearning under general forget requests. Through extensive experiments on Ele-Fashion and Goodreads-NC, we demonstrate that FDQ consistently achieves strong utility preservation while maintaining effective forgetting against membership inference attacks. Overall, FDQ offers a principled and robust solution for privacy-aware unlearning in high-dimensional multimodal graph systems.

Problem

• graph unlearning
• multimodal graphs
• privacy preservation
• utility degradation
• high-dimensional projections

Innovation

• Graph Unlearning
• Multimodal Graph
• Feature-Dimension Awareness
• Quantile Selection
• Privacy-Preserving Learning