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

Researchers propose GDGU, a machine learning technique that enables electric vehicle charging stations to delete training data from deployed cyberattack detection models without full retraining, addressing privacy regulations while maintaining security effectiveness. The method achieves comparable performance to stronger baselines while being 10-12 times faster and more memory-efficient than retraining from scratch.

The intersection of machine learning security and privacy regulation creates significant operational challenges for critical infrastructure operators. Electric vehicle charging networks represent an emerging vulnerability in power distribution systems, requiring sophisticated detection mechanisms to localize cyberattacks before they cascade through feeders. Graph neural networks excel at this spatial-temporal analysis, yet their deployment conflicts with privacy frameworks like GDPR that grant data subjects deletion rights.

Graph unlearning represents a cutting-edge solution to this regulatory-technical tension. Rather than retraining models from scratch—computationally expensive and operationally disruptive—GDGU uses gradient-based approximations to surgically remove individual data points' influence. The technique modifies training data at specific nodes, calculates parameter correction through gradient differentials, and recalibrates batch normalization to restore model utility. Testing across three network sizes (34 to 8,500 nodes) and multiple graph neural network architectures demonstrates robust scalability.

For infrastructure operators and utilities, GDGU enables compliance with privacy obligations without sacrificing cybersecurity posture. The 10-12x speedup over retraining makes data deletion requests operationally feasible rather than prohibitively expensive, reducing incentives to resist privacy demands through implementation barriers. Reduced memory consumption also lowers deployment costs on edge infrastructure alongside charging stations.

The methodology's broader applicability extends beyond EVCS to any graph-structured critical infrastructure requiring both machine learning-driven defense and data governance. As regulations tighten globally, similar unlearning techniques will become essential for power grids, water systems, and telecommunications networks. Performance matching second-order baselines suggests gradient-based approximations can compete with computationally heavier alternatives, potentially reshaping machine learning operations in regulated domains.

• →GDGU enables privacy-compliant data deletion from cyberattack detection models 10-12 times faster than retraining
• →Graph unlearning maintains localization accuracy and security effectiveness while respecting GDPR-like privacy obligations
• →The technique reduces memory requirements compared to second-order alternatives, enabling deployment on edge infrastructure
• →Testing across multiple network topologies and neural network architectures confirms scalability for diverse grid configurations
• →Gradient-based parameter correction offers computationally efficient alternative to full model retraining in regulated industries