Efficient Cyber Attack Detection in Industrial Control Systems Using Lightweight Neural Networks and PCA

Abstract

Industrial control systems (ICSs) are widely used and vital to industry and society. Their failure can have severe impact on both the economy and human life. Hence, these systems have become an attractive target for physical and cyber attacks alike. In this paper, we examine an attack detection method based on simple and lightweight neural networks, namely, 1D convolutional neural networks and autoencoders. We apply these networks to both the time and frequency domains of the data and discuss the pros and cons of each representation approach. The suggested method is evaluated on three popular public datasets, and detection rates matching or exceeding previously published detection results are achieved, while demonstrating a small footprint, short training and detection times, and generality. We also show the effectiveness of PCA, which, given proper data preprocessing and feature selection, can provide high attack detection rates in many settings. Finally, we study the proposed method's robustness against adversarial attacks that exploit inherent blind spots of neural networks to evade detection while achieving their intended physical effect. Our results show that the proposed method is robust to such evasion attacks: in order to evade detection, the attacker is forced to sacrifice the desired physical impact on the system.

Adversarial Attacks on Time-Series Intrusion Detection for Industrial Control Systems

“…Many other models have been analysed. Convolution neural networks [19] have reached high F 1 scores and autoencoders have been investigated in [23], [24] as well as neural architecture search in [25]. Alternatively, ensembles of random trees are used in [26], or a range of classification methods including SVMs, neural networks, and tree based methods are compared in [27].…”

“…Alternatively, generative methods can be used to create adversarial attacks such as in [29] where generative adversarial networks (GANs) were used to create adversarial data. Additional work in [23] investigated adversarial attacks targeting an autoencoder IDS. Unlike [28], the work in [23] modelled their attacker as not having control of the communications to the IDS independently of the programmable logic controller (PLC).…”

“…Additional work in [23] investigated adversarial attacks targeting an autoencoder IDS. Unlike [28], the work in [23] modelled their attacker as not having control of the communications to the IDS independently of the programmable logic controller (PLC). Therefore, the adversarial data had to fool the IDS and fulfill the original cyber-physical attack.…”

“…Having an accurate model of the physical system is a significant requirement, however it is an accepted assumption in this area. Related work provided the attacker with knowledge of all of the data, free of any noise [23], or considered defences which evaluate datapoints independently, and enable attackers to compromise a target without knowledge of the physical system [28]. In comparison our model is more challenging for the attacker.…”

“…The authors of [23] also noticed the discrepancy between the test and training datasets for non-anomalous data. In their work they propose the use of a modified Kolmogorov-Smirnov (KS) test to filter out such features.…”