An intelligent analysis model for detecting anomalies in autonomous vehicle networks with enhanced edge computing

Abstract

Autonomous vehicle in-vehicle networks integrate Controller Area Network (CAN), CAN with Flexible Data-Rate (CAN-FD), in-vehicle Ethernet, domain controllers, and safety-critical electronic control units. Their high communication frequency, strong inter-node coupling, and strict real-time requirements make delayed cloud-based detection unsuitable for protecting perception, braking, steering, and powertrain control links. To address detection delay, uncertain attack propagation, and limited adaptability to edge deployment, an intelligent intrusion analysis model for autonomous vehicle networks is developed. A convolutional neural network–bidirectional gated recurrent unit (CNN-BiGRU) architecture learns local message disturbances and temporal communication dependencies, while an attention mechanism highlights high-risk windows for identifying denial-of-service (DoS), Fuzzy, Replay, and Spoofing attacks. The recognition output is further mapped to the in-vehicle topology by combining node criticality and link association strength to quantify propagation risk and determine response priority. Simulation results show an accuracy of 97.38%, an F1-score of 96.89%, and an inference latency of 14.2 ms under 8 tera operations per second (TOPS). The model maintains stable detection under different edge computing capacities and connects intrusion recognition, propagation assessment, and local response into a continuous evaluation chain, providing millisecond-level security support for safety-critical communication and control in autonomous vehicles and reducing the time available for malicious traffic to affect driving decisions.