CGAN-Augmented Federated Learning for Adaptive Multi-Strategy Jamming Detection in Stochastic Tactical Air–Ground Network

Abstract

Tactical wireless networks, particularly stochastic tactical air-ground networks (STAGNs), operate in highly adversarial and dynamic environments. Accordingly, these systems face increasing challenges from adaptive multi-strategy jamming attacks that closely resemble legitimate signals and disrupt communication reliability. In this context, we consider adaptive multi-strategy jamming (AMSJ), modeled as a stateless multi-armed bandit (MAB) that dynamically selects its interference strategy according to instantaneous SINR degradation feedback. The jammer alternates among bandwidth fragmentation, frequency sweeping, intermittent power bursts, and adaptive narrowband jamming. This adaptive behavior generates non-stationary interference patterns that severely degrade signal quality and make jammed signals difficult to distinguish from unjammed ones. To address these challenges, in this study, we propose Fed-AugCGAN, a privacy-preserving federated learning (FL) framework based on federated augmented conditional GANs. The method applies band-pass filtering and short-time Fourier transform (STFT) spectrograms to extract discriminative jamming patterns, while each UAV trains a local CGAN to generate class-conditioned synthetic samples and mitigate local data imbalance without sharing raw signals. Then, the server aggregates discriminator updates using a validation-guided weighted aggregation that increases the contribution of more reliable local models under heterogeneous non-IID conditions. The proposed approach improves detection accuracy and robustness against AMSJ attacks while remaining scalable to STAGNs.