GAN Neural Networks Architectures for Testing Process Control Industrial Network Against Cyber-Attacks

Abstract

Protection of computer systems and networks against malicious attacks is particularly important in industrial networked control systems. A successful cyber-attack may cause significant economic losses or even destruction of controlled processes. Therefore, it is necessary to test the vulnerability of process control industrial networks against possible cyber-attacks. Three approaches employing Generative Adversarial Networks (GANs) to generate fake Modbus frames have been proposed in this work, tested for an industrial process control network and compared with the classical approach known from the literature. In the first approach, one GAN generates one byte of a message frame. In the next two approaches, expert knowledge about frame structure is used to generate a part of a message frame, while the remaining parts are generated using single or multiple GANs. The classical single-GAN approach is the worst one. The proposed one-GAN-per-byte approach generates significantly more correct message frames than the classical method. Moreover, all the generated fake frames have been correct in two of the proposed approaches, i.e., single GAN for selected bytes and multiple GANs for selected bytes methods. Finally, we describe the effect of cyber-attacks on the operation of the controlled process.

Keywords

• GAN neural networks
• cyber-security
• cyber-attacks
• industrial network