NodeVanisher: Efficient and Scalable Data Unlearning in GNNs through Targeted Subgraph Refinement

Citations

SCOPUS

0

초록

Unlearning specific data in Graph Neural Networks(GNNs) is crucial for ensuring data privacy and enabling modeladaptability. Existing graph-unlearning methods face a trade-offamong computational cost, fidelity to full retraining, and formalguarantees when removing specific data. To address these limitations, this paper presents NodeVanisher, an efficient and scalableframework that enables localized unlearning in GNNs. It employsa targeted subgraph refinement strategy combined with dynamicneighbor sampling to enable selective unlearning of specificnodes without full model retraining. The proposed method improves efficiency and minimizes performance loss by preservingstructural relationships using subgraph refinement. Experimentalevaluations on multiple datasets demonstrate that this approachsignificantly reduces retraining time while maintaining accuracycompetitive with full retraining and with recent unlearningbaselines. We further provide empirical privacy evidence withmembership inference attacks (MIA), demonstrating reduced information leakage after unlearning across all GNN architectures.NodeVanisher reduces the localized retraining time by up to104× (vs. the full-retraining baseline on the remaining graph) onPhysics, 22× on PubMed, and 30× on Citeseer. Additionally, itachieves gains of +10.7 points on Citeseer/GCN and +3.2 points onPhysics/GIN relative to reproduced strong unlearning baselines(e.g., GNNDelete). Furthermore, the experimental results presentthe potential of our approach as a robust, scalable, and privacyenhancing solution for efficient localized data unlearning in dynamic graph environments.

키워드

Graph Neural Networks; Graph Unlearning; Node Unlearning; Privacy; Privacy-Preserving AI; Security
제목
NodeVanisher: Efficient and Scalable Data Unlearning in GNNs through Targeted Subgraph Refinement
저자
Ahsan, Imran; Gupta, Brij B.; Kim, Jinsung; Kim, Mucheol
DOI
10.1109/TCE.2026.3726041
발행일
2026
유형
Article
저널명
IEEE Transactions on Consumer Electronics