Privacy-Preserving Anomaly Detection Using Vertical Federated Learning in Smart Grids

The increasing digitalization of power systems has heightened the risk of cyber-physical attacks on critical infrastructure. However, traditional centralized anomaly detection approaches often require full data visibility, which conflicts with privacy and regulatory constraints in operational environments. This paper presents a Vertical Federated Learning (VFL) based framework for unsupervised anomaly detection in power systems, enabling collaborative training without sharing raw data across clients. A fused cyber-physical dataset [13], incorporating Snort alerts and DNP3 telemetry from a simulated IEEE 9-bus system, was vertically partitioned to reflect realistic feature distributions across substations. Two unsupervised models, Autoencoder and One-Class Support Vector Machine (OCSVM) were implemented and evaluated under both centralized and VFL configurations. Experimental results demonstrate that the VFL OCSVM an F1 score of 85%, closely approaching the centralized model’s performance of 87%, while ensuring data privacy. The VFL Autoencoder also showed reasonable accuracy, albeit with slightly reduced performance due to partial feature visibility. Comparative analysis highlights the effectiveness of OCSVM in distributed anomaly detection settings and confirms that VFL introduces only marginal performance degradation compared to centralized baselines