GDGU: A Gradient Difference-based Graph Unlearning Method for Cyberattack Localization in Electric Vehicle Charging Networks

The article presents GDGU, a Gradient Difference-based Graph Unlearning method designed for cyberattack localization in Electric Vehicle Charging Networks. This approach formulates unlearning as a feature-level problem within a multi-label classification task on graph data, utilizing a first-order parameter correction based on gradient differences. Benchmarked against second-order unlearning methods on various distribution networks, GDGU demonstrates comparable localization utility while being 10 to 12 times faster and more memory-efficient than full retraining, which is crucial for practitioners needing efficient model updates in compliance with privacy regulations.