Adaptive multi-channel topology-augmented graph neural networks for secure node classification in API-based power systems

Abstract Graph neural network models have demonstrated strong effectiveness across a wide range of graph-oriented learning problems. In complex cyber-physical environments like smart grids, component interactions and multi-domain data exchanges are typically handled through application programming interfaces (APIs), resulting in intricate interaction graph structures with severe network security and asset protection considerations. Within these cloud-native and edge-integrated power API environments, the empirically constructed graph topology is often highly sparse, noisy, or partially missing due to telemetry dropouts or gateway blind spots, thereby reducing the node classification performance of conventional GNN approaches. To address these industrial challenges, we introduce an adaptive multi-channel topology-enhanced graph neural network (AMTGNN), which improves structural information via multiple augmentation techniques and integrates them through adaptive fusion. Specifically, three complementary topology enhancement strategies are developed to match power system semantics, including random walk-based, K-nearest neighbor-based, and adaptive threshold-driven augmentation methods. To combine the information from different augmented views, a channel-level attention mechanism is employed to automatically assign importance weights to each channel based on local structural characteristics. Furthermore, a topology-aware loss term is incorporated to promote representation consistency among connected nodes. Extensive evaluations on representative benchmarks verifying structural equivalents demonstrate the superiority of AMT-GNN for secure endpoint classification, while ablation studies further verify the contributions of the topology augmentation module, attention mechanism, and topologyaware loss to overall cyber-security defense performance gains.