DRL based DDQN for smart jamming mitigation in massive MIMO uplink systems
Аннотация
Time division duplexing (TDD) massive multiple-input multiple-output (MIMO) systems depend on uplink pilot training to acquire channel state information (CSI). However, reusing non-orthogonal pilot sequences creates pilot contamination, which corrupts minimum mean square error (MMSE) channel estimates. Although this vulnerability was historically considered a benign multi-cell interference issue, it is now being weaponized by intelligent jammers using hybrid tactics, such as pilot sweeping and block-delayed reactive imitation, to cause severe beamforming misalignment and degrade signal-to-interference-plus-noise ratios (SINR). Since traditional static or random-hopping countermeasures are ineffective against adaptive adversaries, this paper introduces a model-free deep reinforcement learning (DRL) defense. Specifically, we deploy a double deep Q-network (DDQN) agent at the base station that operates within the physical-layer MMSE estimation framework. Driven by real-time, post-combining SINR feedback, the agent autonomously formulates an optimal evasive pilot allocation strategy without requiring extensive multi-cell coordination. Monte Carlo simulations across 100 independent seeds confirm that the DDQN framework effectively mitigates adversarial actions. It reduces pilot collisions, recovers spectral efficiency, and restores the empirical bit error rate (BER) to secure, unjammed levels.