On-Board Machine Learning for Satellite Cybersecurity

Objectives

1. Explore state-of-the-art in multiagent collaboration and identify open problems.
2. Develop novel frameworks that combine knowledge-based and data-driven methods to address these open problems.
3. Implement and evaluate the algorithms in the context of challenging simulated and/or real-world environments and tasks.

Description

Satellite systems are increasingly exposed to cyber threats as reliance on space-based services grows across civil, commercial and defence sectors. Although current frameworks emphasise the importance of on-board intrusion detection, practical guidance and empirical research on approaches for on-satellite cybersecurity remain limited. The variability in the nature of threats and the wide range of factors to be considered motivate the use of Machine Learning (ML) methods. However, existing literature on machine learning (ML)-based anomaly and intrusion detection for satellite systems has predominantly focused on ground-based anomaly detection and telemetry analysis, with very limited work on an ML-enabled cybersecurity model deployed on-board.  In addition, modern ML methods that have been applied to telemetry or command-and-control data are based on deep network models such as CNNs, LSTMs, and autoencoders. These methods and models have to rely on off-board processing and it will be difficult to use them for on-board implementation due to resource constraints, limited datasets, and the absence of in-orbit testing.   Overall, there are significant gaps in empirical validation, real-time on-orbit operation, and the integration of fault detection with cybersecurity objectives. This doctoral research project seeks to address these gaps by exploring lightweight, resource-efficient, on-board ML systems based on principles of multiagent reasoning and learning, which are capable of automated threat detection and response, as well as subsystem-specific optimisation for enhanced mission resilience.