Deep Learning and Explainable AI Models for Intrusion Detection in Space-Ground Communication Networks: A Review.

Abstract:

This review critically evaluates the suitability of deep learning and explainable artificial intelligence approaches for intrusion detection in satellite ground-station environments, addressing the escalating cybersecurity risks facing the National Space Research and Development Agency (NASRDA) and broader space communication networks. Using a systematic narrative review across IEEE Xplore, ACM, Scopus and arXiv, the analysis compares CNN, LSTM, GRU, autoencoder and transformer-based IDS models, revealing that while reported accuracies frequently exceed 92% on benchmark datasets, performance declines by 20% to 35% under domain shift, demonstrating poor transferability to space–ground telemetry. XAI methods such as SHAP, LIME and Integrated Gradients appear in more than 80% of reviewed studies, yet empirical results show a 30% to 60% increase in inference latency, raising concerns about operational feasibility in real-time satellite control systems. A mathematical hybrid model combining CNN, LSTM and transformer components with a structured anomaly-scoring function and explanation regulariser is formulated to address these limitations. Findings indicate that multi-model fusion enhances anomaly sensitivity, domain-specific feature engineering improves robustness, and integrated XAI pathways strengthen analyst trust while exposing computational bottlenecks. The proposed conceptual architecture for NASRDA advances the field by aligning detection workflows, interpretability mechanisms and feedback loops with the constraints of aerospace communication systems. The review concludes by identifying key research priorities, including the development of satellite-specific datasets, real-traffic validation of hybrid IDS models, and deployment of low-latency XAI dashboards for operational security.