A Real-Time Temporal Deep Learning Framework for Anomaly Detection in Autonomous Vehicles
Raed Hameed, P. Nagarathna, Bhagyashree Kulkarni, S. Senthil Kumar, S. P. R. Swamy Polisetty · 2025
In recent years, real-time cyberattack detection in autonomous vehicles has become critical for ensuring safe and reliable operation. However, existing models are trained on simulated data using static image-based classification and lack temporal awareness and generalizability to real-world deployment. The proposed Real-Time Controller Area Network (CAN)-based Anomaly Detection (RT-CANAD) framework leverages temporal deep learning on CAN data streams. Initially, raw CAN signal data is preprocessed through normalization and time windowing to ensure clean and structured input. Furthermore, to extract local patterns and short-term dependencies from CAN signal sequences, a hybrid 1D-CNN is used. Then, Gated Recurrent Unit (GRU) captures sequential and temporal dependencies in time series to understand long-term behavior. Moreover, it also allows accurate detection of cyber threats, such as spoofing, denial-ofservice, and injection attacks. Real-world CAN intrusion datasets are used to train and evaluate model. Furthermore, Tensor RT is used to optimize and deploy the trained model on embedded edge devices, allowing sub-100ms inference latency. The, experimental results show that proposed system achieves accuracy (98.6%), precision (96.4%), recall (95.8%), and F1score (99.1%) while maintaining high inference efficiency, making it suitable for real-time protection in autonomous vehicle networks.