Deep learning–based anti-jamming detection and suppression for communication signals in complex electromagnetic environments

Abstract

Communication links operating in complex electromagnetic environments are vulnerable to broadband noise, frequency-sweep interference, and carrier-modulated deceptive jamming, especially under low signal-to-noise ratio (SNR) conditions. This work proposes a unified anti-jamming system that jointly performs interference classification and spectral suppression/reconstruction from short-time Fourier transform (STFT) representations. The model integrates a convolutional neural network (CNN) backbone for local time – frequency texture extraction, a bidirectional long short-term memory (LSTM) for temporal dependency modeling, and a self-attention module to emphasize interference-dominant regions. To enhance robustness to non-stationary jammers, we apply channel-noise augmentation over an SNR range of −8 dB to 6 dB during training. Experiments on a composite dataset covering satellite, terrestrial, and indoor simulated links demonstrate stable performance under strong interference. The main contribution lies in the joint multi-task learning formulation that simultaneously performs interference classification and spectral suppression within a unified CNN-BiLSTM-Attention framework. Practically, the system achieves an average SIR improvement of 14 dB and maintains BER within 10⁻³ under low - SNR conditions (≥8 dB to 6 dB), with an end-to-end latency of 15 ms on a GPU platform, making it suitable for real-time UAV command links and satellite downlinks. After suppression, the recovered signal quality supports demodulation under low-SNR inputs, and the measured BER remains within the 10⁻³ order (i.e., BER ≤ 10⁻³ under most tested conditions). The end-to-end inference latency is approximately 15 ms on a GPU platform, satisfying real-time processing constraints for short-frame communication enhancement.