Federated Spatiotemporal Graph Learning for Passive Attack Detection in Smart Grids

Researchers present a federated learning approach to detect passive eavesdropping attacks in smart grids by combining graph neural networks with temporal modeling. The system achieves 98.32% per-timestep accuracy while preserving data privacy through decentralized training, addressing a critical vulnerability in grid infrastructure where attackers silently gather topology and consumption data.

Smart grid security faces an underappreciated threat in passive reconnaissance attacks, where adversaries extract sensitive operational intelligence without altering data. This research tackles the detection challenge through a novel spatio-temporal architecture that identifies subtle electromagnetic signatures across distributed network nodes simultaneously. The approach proves particularly valuable because passive attacks produce faint, transient signals invisible to single-node monitoring—a weakness that federated graph learning uniquely addresses by aggregating context across ego-centric network neighborhoods while maintaining temporal coherence through bidirectional GRU layers.

The technical contribution leverages federated learning under FedProx, enabling critical infrastructure operators to train robust detection models without centralizing raw meter data. This addresses both privacy concerns and the practical reality of heterogeneous smart grid deployments spanning home area networks (HAN), neighborhood area networks (NAN), and wide area networks (WAN). The synthetic dataset generation methodology, informed by grid communication standards, creates realistic non-IID training conditions that enhance real-world applicability.

For grid operators and security practitioners, the 98.32% per-timestep accuracy with 0.15% false-positive rates represents a deployable solution rather than theoretical research. The decision rule using run-length filtering (m=2) and threshold tuning (τ=0.55) enables practical implementation without excessive operational friction. The work's significance extends beyond energy infrastructure—the spatio-temporal federated learning framework demonstrates how privacy-preserving anomaly detection applies across critical infrastructure sectors from water systems to transportation networks. Future deployment depends on validating performance against real-world passive attacks and integrating results into operational grid management workflows.

• →Federated graph neural networks detect passive smart grid eavesdropping with 98.32% accuracy while preserving data privacy on client devices.
• →Combining spatial graph convolution with temporal GRU modeling identifies stealthy reconnaissance signals invisible to single-node monitoring.
• →The approach maintains 0.15% false-positive rates, making it operationally viable for real critical infrastructure deployment.
• →Non-IID federated learning handles heterogeneous HAN/NAN/WAN network architectures typical of distributed smart grid environments.
• →The methodology establishes a template for privacy-preserving anomaly detection applicable across other critical infrastructure sectors.