Module ablation results.

With the rapid development of smart grids and the Power Internet of Things (PIoT), wireless communication networks are facing the severe threat of dynamic eavesdropping attacks. Traditional detection methods rely on static assumptions or shallow models, which are not capable of dealing with complex topology mutations and high-dimensional nonlinear features. There is an urgent need for efficient and lightweight adaptive solutions. This study proposes a Dynamic Spatiotemporal Fusion Framework (DSTF-GKAN), which integrates the spatiotemporal dynamic modeling capability of Graph Recurrent Neural Networks (GRNN) with the lightweight adaptive spline approximation mechanism of Kolmogorov-Arnold Networks (KAN). By adaptively optimizing the mesh to dynamically adjust the spline control points and introducing hierarchical sparse regularization to compress parameters, the model enhances its sensitivity to channel anomalies through the integration of physical layer security (PLS) feature constraints. Experimental results show that under dynamic scenarios with an attack mutation rate (AMR = 0.5), DSTF-GKAN achieves a detection F1 score of 0.891, which is a 7.1% improvement over GRNN, and reduces the localization error (RMSE = 0.518 m) by 16.2%. After quantization and pruning optimization, the model has a parameter size of only 0.2 MB, with an inference latency of 0.9 ms and energy consumption of 16mJ on edge devices. Ablation experiments have verified the necessity of the GRU-GCN module (contributing 4.9% to the F1 score) and PLS regularization (improving the F1 score by 1.3%). DSTF-GKAN provides an efficient, robust, and interpretable detection framework for smart grid security. Its lightweight design promotes real-time edge defense and lays the theoretical and technical foundation for the construction of a secure energy internet ecosystem.