This study explores the critical area of cyberattack detection and diagnosis in the scope of offshore control systems, specifically focusing on Modbus TCP/IP communication between refrigeration systems and SCADA. The project emphasizes the importance of swiftly and accurately identifying cyberattacks, followed by predicting system responses using neural networks.
The research incorporates the design and implementation of five distinct Long Short-Term Memory (LSTM) neural networks (NNs), each independently tailored for the classification or regression of various operational and cyberattack scenarios based on the respective system reaction. The first three LSTM NNs focus on the classification of 4 cyberattacks, while the other two forecast the room temperature following 5 hours into an attack detection. The uniqueness of each NN lies in their use of different solvers, architecture and hyperparameters.
The use of classification and forecasting models offers a comprehensive approach to handle cyber threats. While the classification networks effectively classified 100 datasets into their respective healthy or cyberattack scenario, the regression networks predicted the room temperature after a cyberattack detection 5 hours in advance with a RMSE of 0.17 and 0.19.
While providing promising results, it also opens doors to further enhancements, potentially contributing to the development of a more robust and effective cybersecurity framework for offshore industrial control systems.