Electric vehicle charging stations (EVCSs) are increasingly vulnerable to cyberattacks that can compromise distribution feeders. Machine learning models, particularly graph neural networks (GNNs), can localize which bus is attacked, but privacy regulations granting EVCS owners the right to delete their training data from deployed models create a computational bottleneck: retraining from scratch for every deletion request is prohibitive, according to a new paper on arXiv.
The Data Deletion Problem in Cyberattack Localization
As EVCSs become more interconnected, the need for accurate cyberattack localization grows. However, data privacy laws (such as the right to erasure) require that upon request, an owner's data must be removed from any trained model. Full retraining after each deletion is computationally expensive, especially for large distribution networks. The paper, authored by Liu, Nanhong; Sun, Mucun; and Zhang, Jie, tackles this by applying graph unlearning (GU) — a technique that efficiently removes the influence of specific data without starting over.
GDGU: Gradient Difference-Based Graph Unlearning
The proposed method, called gradient difference-based graph unlearning (GDGU), formulates the problem as a feature-level unlearning on a graph-level multi-label classification task. It removes the influence of requested deletion data through a first-order parameter correction, computed from the gradient difference between the original training data and a modified dataset where only the charging power features at the requested EVCS buses are unlearned. After correction, the method applies batch-normalization recalibration and a brief recovery fine-tuning step to restore localization utility.
Benchmark Performance on IEEE Distribution Networks
The researchers benchmarked GDGU against two second-order GU baselines on three IEEE distribution network test cases (34-bus, 123-bus, and 8500-node) using three different GNN backbones under cumulative unlearning scenarios. The results are summarized below:
| Metric | GDGU Performance | Comparison Baseline |
|---|---|---|
| Localization utility | Matches strongest baseline | Second-order GU baselines |
| Forgetting fidelity | Close to full retraining | Full retraining is baseline |
| Unlearning speed | 10 to 12 times faster | Retraining from scratch |
| Memory usage | Far less than baselines | Second-order GU baselines |
GDGU achieves forgetting fidelity comparable to full retraining while being an order of magnitude faster and using significantly less memory, according to the paper.
Implications for EV Charging Infrastructure Cybersecurity
The ability to efficiently remove user data from deployed models without sacrificing attack localization accuracy is critical for both compliance and operational security. As EV charging networks expand, operators will need tools that respect data rights while maintaining robust defenses. GDGU offers a practical path forward, balancing privacy, speed, and accuracy. The method's reliance on first-order corrections rather than computationally heavy second-order approximations makes it particularly suitable for real-world deployment where memory and time are constrained.
