Heterogeneous Federated Deep Reinforcement Learning-Empowered Dual-Threat Jamming Detection in Space–Air–Ground ISAC Networks

In this paper, we address the challenges faced by space-air-ground integrated sensing and communication (SAG-ISAC) networks, where the coexistence of sensing and communication functionalities is disrupted by dual-threat jammers, leading to severe degradation in network performance and security. The situation is further complicated by the heterogeneous nature of the collected data and non-independent and identically distributed (non-IID) patterns, which make it challenging to apply traditional machine learning approaches. Furthermore, protecting data privacy in distributed environments adds a critical layer of complexity, which requires a paradigm that balances robustness, adaptability, and security. To tackle these issues, we propose a resilient federated deep reinforcement learning (ReFDRL) framework for detecting dual jammers in SAG-ISAC networks by integrating tailored Vision Transformer (ViT) models and deep Q-learning (DQL) for dynamic weight aggregation. The proposed solution introduces ViT models for sensing and communication devices, with UAVs utilizing time-frequency distribution (TFD) features for sensing signals, while user equipments (UEs) employ spectral correlation function (SCF) features for communication signals. The proposed framework combines federated learning (FL) for privacy-preserving distributed training with a DQL-based aggregation mechanism at the aggregator to dynamically adjust the contribution of local model updates according to their performance. Accordingly, the proposed solution manages distributed heterogeneous data through FL while adaptively balancing sensing and communication contributions under varying jamming conditions. Numerical results demonstrate that the proposed ReFDRL framework improves the detection of dualthreat jammers, effectively handles non-IID data, and achieves a reliable trade-off between sensing and communication performance in SAG-ISAC networks.