基于指数平滑动态图的CAN总线入侵检测方法OA

The security of controller area network(CAN)bus is increasingly challenged by volatile and non-stationary communication patterns in modern vehicles,which traditional static detection methods have failed to capture.ES-DyGNN,an exponentially smoothed dynamic graph neural network,was proposed to capture the evolving relationships between electronic control unit(ECU).Unlike heuristic dynamic models,this method was underpinned by a rigorous ex-ponential smoothing graph operator that adaptively captured topological shifts.Closed form expansions for the dynamic adjacency sequences were derived and Frobenius norm convergence bounds that characterized the stability of the model were established.Furthermore,a theoretical lower bound on attack persistence was proven,ensuring subtle injections were detectable despite noise.Additionally,the model employed sinusoidal time embeddings and edge-conditional atten-tion to weigh both feature similarity and transition frequencies during message passing.Through extensive evaluations on benchmark datasets,it was demonstrated that an accuracy of over 99%was achieved by ES-DyGNN,while an infer-ence latency of less than 0.14 ms for each window was sustained.Through both rigorous theoretical analysis and exten-sive experimental validation,the proposed method demonstrates the feasibility of topology adaptation for automotive se-curity.