Transferrable Graphon-Based Detection of Adversarial Samples in Power Grid Networks

Modern grid infrastructures rely heavily on connected devices and real-time communication for efficient and secure operation. While traditional machine learning, deep learning, and graph-based methods have shown success in detecting classic false data injection attacks (FDIAs), which manipulate readings using simple perturbation strategies, their performance degrades significantly against evasion attacks where adversaries generate adversarial samples to fool detection models without requiring system knowledge. Moreover, traditional graph-based approaches suffer from high computational overhead when trained on large power system topologies. In this work, we propose implementing a graphon neural network (WNN)-based detector against evasion attacks, assuming a realistic threat scenario where the adversary has no prior system knowledge. Unlike graph neural networks, which require retraining when the topology changes, WNNs enable efficient and transferable learning by modeling families of graphs that share structural properties. Experimental results demonstrate that our WNN-based detector maintains robust detection against evasion attacks with a 1–2% deterioration in detection rate, compared to a 9–10% deterioration in benchmark models, while enhancing decision-making latency by 116% and reducing training time by 133% compared to benchmarks.