Novel Multi-Stage Multi-Task Deep Learning Framework for the Cybersecurity of In-Vehicle Networks

Objective - To develop a deep learning-based robust threat detection system for in-vehicle networks of autonomous vehicles to differentiate normal vehicle operation from potential cyber-attacks.Method - The Controller Area Network (CAN) network traffic dataset, including both attack and normal messages, was used for the study. Several features were present in this dataset, including class and subclass designations that help to differentiate between legitimate operational signals and possible attack messages. Feature engineering techniques were applied to the dataset. Data preparation was done after feature engineering, with the main goal being to make it ready to train the model.After pre-processing the data, three distinct neural network architectures were applied utilizing a multi-task and multi-stage approach: convolutional neural network, long short-term memory, and artificial neural network. The most appropriate and useful model for the specific context was then identified by comparing the accuracies of different models.Result - With a multistage approach, the convolutional neural network model demonstrated remarkable training accuracies of 98.10% for Stage I (classifying attack or normal condition) and 99.08% for Stage II (classifying type of attack). Also, the Stage I and Stage II testing accuracies were 98.08% and 99.09%