Securing Industrial Control Systems from False Data Injection Attacks with Convolutional Neural Networks

Abstract

Due to trends in modern infrastructure development and usage, the attacks on Industrial Control Systems (ICS) are inevitable. New threats and other forms of attacks are constantly emerging to exploit vulnerabilities in system compromising the security parameters such as Confidentiality, Integrity and Availability (CIA). Injection attacks also termed as False Data Injection Attacks (FDIA) are the complex attacks on the ICS. FDIA affects the data integrity of a packet by modifying their payloads and are considered as an intrusion via remote access. In FDIA, attackers gain access to a critical process or process parameters in ICS and forces them to execute according to the newly injected code or command. For our research, a process control plant from Integrated Automation laboratory was used to acquire different parameters related to ICS. Injection attacks such as measurement injection and command injection were simulated and injected into the obtained plant data. Convolutional Neural Networks (CNN) is used to evaluate the functionality of identifying those injection attacks. Multiple steps such as pre-processing, feature extraction, data transformation and image representation were performed in order to feed the CNN with the simulated plant data. A 3-layered fully connected CNN architecture with non-linear ReLU activation is built along with a SoftMax classification layer for classifying the input data as a normal or an attack. A proper training of CNN is done by checking the variance to avoid overfitting and underfitting of the network. Performance parameters such as accuracy, recall, precision F-measure and Cohen’s kappa coefficient were computed. CNN outperforms in the performance compared to other deep learning approaches.