Inproceedings,
Intelligent sensor attack detection and identification for automotive cyber-physical systems
2017 IEEE Symposium Series on Computational Intelligence (SSCI), page 1-8. (November 2017)
DOI: 10.1109/SSCI.2017.8280915
Abstract
This paper addresses the problem of detection and identification of the sensor attacks when most sensors are attacked. Sensors can play a key role to improve safety and convenience in automotive Cyber-Physical Systems (CPS). A dramatic increase in connectivity and openness of the automotive CPS brings high security risks. If multiple and heterogeneous sensors equipped for braking and steering provides false sensing information for their controllers under deception attacks, it might cause catastrophic situations during driving. If the existing machine learning approaches are applied for sensor attacks while the majority of sensors is attacked, it cannot guarantee to identify deceptions as cyber-physical attacks. To address this problem, we propose an intelligent sensor attack detection and identification method based on Deep Neural Network (DNN) techniques, called deep learning, without a prior knowledge about the deception attacks modifying sensing data in time. We investigate an autonomous vehicle with Inertial Measurement Unit (IMU) and wheel encoder sensors under conditions of uncertainty and nonlinearity during driving. We firstly identify all possible attacks category on the sensors of it, choose what model to use and then systematically design its architecture on which the performance of deep learning highly depends. We train and then validate the proposed method's performance on real measurement data obtained from an unmanned ground vehicle. Finally, we show analytically the superiority of our method in terms of accuracy, precision, and computation time, including the worst situation where two among three sensors are simultaneously attacked.
