Adaptive Resilience and Self-Healing in Aerospace System-of-Systems (SoS) Using Graph Neural Networks and Adversarial Machine Learning

Complex aerospace systems of systems (SoS), like UAV networks, satellite constellations, and sensor systems, work in highly dynamic and adversarial environments. This research deals with building a self-healing and adaptive system within this complex system (SoS) using Graph Neural Networks (GNNs) and Adversarial Machine Learning (AML) that can predict, detect, and then mitigate adversarial attacks and actively manage real-time faults. Our idea was to devise a multi-layered defense framework that not only deals with anomalies caused by adversarial attacks like spoofing, jamming, data poisoning, etc. but also predicts fault propagation and ensures robust decision-making even under adversarial conditions across the system before some major critical failures occur in aerospace missions. We introduce an innovative framework that involves fault diagnosis and prognosis (FDP) techniques along with graph neural networks (GNNs) and adversarial machine learning (AML) to improve resilience, proactive maintenance, and anomaly detection in SoS. In this paper, we tried to demonstrate how the GNNs are applied to model inter-agent dependencies and cope with the faults in real time while the AML techniques help the system to identify, learn, and defend itself against adversarial attacks. We compared multiple fault detection techniques, including K-means clustering, autoencoders, and GNNs (with and without adversarial training). The results reveal that the regular GNN and GNN-AML frameworks improve fault detection accuracy by 80% and UAV swarm stability under adversarial situations, resulting in a 22 % increase in system resilience. This technique provides a strong foundation for future aircraft system operations in dynamic and potentially dangerous circumstances.