MLIDS: Handling Raw High-Dimensional CAN Bus Data Using Long Short-Term Memory Networks for Intrusion Detection in In-Vehicle Networks

Abstract

CAN uses no authentication and encryption mechanisms for secure communication. To solve the security issues of the CAN bus, a deep learning-based intrusion detection systems have been proposed. But due to the high dimensional property of the CAN bus data, it was not possible to create an effective Intrusion Detection System (IDS) in the CAN bus that can take the property of the CAN data into consideration. In this paper, we are proposing a Long Short-Term Memory Networks (LSTM) based IDS that can handle the high dimensional property of the CAN bus data . Unlike the conventional methods which required a single network architecture for each unique arbitration ID, our method gives a single overall anomaly signal over a certain detection window without the need for reverese-engineering the CAN bus data. Using this anomaly signal we have managed to achieve 100% detection precision for insertion, fuzzy and targeted attacks in our data and in a public data that is prepared for this specific purpose.