DRL-Based Secure Adaptive Waveform for Satellite Communications

Abstract

Conventional modulation techniques using fixed constellations, such as QAM and PSK, are increasingly challenged in dynamic and non-stationary satellite channels due to their limited adaptability and vulnerability to interception and signal classification. This paper proposes a deep reinforcement learning (DRL)-based adaptive waveform framework in which each symbol is represented by a short complex-valued waveform rather than a single constellation point. Unlike fixed QAM/PSK schemes, the proposed framework jointly adapts modulation order and waveform geometry to instantaneous SNR variations by decoupling discrete and continuous decisions: a DQN selects the modulation order from an SNR-dependent finite set, while a DDPG agent optimizes the waveform structure. The resulting multi-sample waveform symbols reduce sensitivity to channel non-stationarity (e.g., LoS/NLoS switching) and inherently degrade blind modulation classification by disrupting symbol-bit correspondences, achieving a balanced reliability-security trade-off at the cost of a 2.3 dB BER increase relative to ideal fixed-constellation modulation.