LKG-RiskNet: legal knowledge-guided graph neural networks for data security risk prediction in cyber-physical power systems

Abstract Data generated by cyber-physical infrastructures—such as smart grids, renewable-energy plants, electric-vehicle (EV) platforms, and industrial IoT—has become a core operational asset. The continuous generation of sensitive telemetry and operational logs makes data-security violation risk not just a legal-compliance issue, but a critical electrical-engineering and system-security challenge. Existing assessment methods, often relying on manual rule checking or shallow statistical learning, struggle to jointly capture legal semantics, enterprise relations, and engineering data flows. To address this, we propose LKG-RiskNet, a legal knowledge-guided graph learning framework for predicting corporate data security violation risk. First, we construct a legal-engineering knowledge graph linking regulations, compliance obligations, penalty criteria, and diverse engineering scenarios. Second, we build an enterprise risk graph integrating penalty cases, business registration, and electrical-engineering attributes. Third, we design a cross-graph fusion mechanism that injects legal and engineering knowledge into enterprise representation learning via compliance-aware relation attention, rule-path aggregation, and rule-consistency regularization. The framework outputs risk probabilities and provides interpretable evidence by tracing influential legal-engineering paths. Experimental results on the real-world DSV-EE-CN dataset demonstrate that LKG-RiskNet achieves an AUC of 0.915 and an AUPR of 0.627, significantly outperforming state-of-the-art graph baselines under class imbalance and sparse-relation settings. This study advances intelligent compliance warning by bridging symbolic legal knowledge and graph neural reasoning in power-system applications.