SpIDER: Space Satellite Intrusion Detection Using Explainable Reinforcement Learning for Next-Generation Satellite Communication Networks

Space and satellite-based systems have had a monumental impact on providing greater interconnectivity across the world. The usage of space and satellite-based systems has increased the ability to access internet resources even in remote areas. Unfortunately, these systems are subject to malicious and multi-faceted cyberattacks. Therefore, proper threat detection systems must be implemented to safeguard these space systems. In our study, we present our novel intrusion detection framework, SpIDER, a space satellite intrusion detection system using explainable reinforcement learning. SpIDER leverages the benefits offered by reinforcement learning and Shapley additive global explanations to improve both the performance and explainability of space-based intrusion detection. We compare our SpIDER framework to several popular machine learning algorithms using the STIN and NSL-KDD datasets. We observe that our SpIDER framework achieves high performance, with accuracy and G-Mean above 99.98% on the STIN satellite dataset. SpIDER also outperforms other machine learning models on the NSL-KDD local area network dataset, achieving accuracy of 76.71% and a G-Mean of 80.49%. These results demonstrate that our SpIDER explainable deep reinforcement learning framework can perform as well or better than supervised machine learning models on both satellite-style and local area network data.