Robust <i>μ</i> -Channel Estimation for IoT and 6G Edge Devices via Defensive Distillation

Reliable channel state information (CSI) is a critical enabler for low-power Internet of Things (IoT) links and emerging 6G edge devices, where receivers must operate under tight energy/latency budgets and in the presence of non-ideal noise and malicious interference. Deep learning (DL)-based channel estimators can surpass classical LS/MMSE baselines; however, they remain vulnerable to distribution shifts and adversarial attacks targeting pilot observations. This paper proposes a lightweight and robust micro-channel estimation (μ-CE) framework based ondefensive distillation, where a compact student convolutional neural network (CNN) is trained under a higher-capacity teacher estimator using a regression-oriented distillation loss. The resulting μ-CE learns a smoother input–output mapping with reduced gradient sensitivity, improving trustworthiness without sacrificing accuracy or computational efficiency. Using MATLAB-and DeepMIMO-generated 5G NR TDL-C channels, we evaluate robustness under diverse non-adversarial noise types and four white-box gradient-based attacks (Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), Momentum Iterative Method (MIM), and Projected Gradient Descent (PGD)). Compared with an undefended CNN, the proposed μ-CE improves normalized mean squared error (NMSE) by approximately 0.5–1 dB under the considered non-adversarial noise conditions (including additive white Gaussian noise (AWGN) at a signal-to-noise ratio (SNR) of 15 dB in the default test setting), limits adversarial NMSE degradation to within 1–2 dB of the clean baseline for moderate perturbation budgets, and reduces attack success rate (ASR) by about 25–40%. Moreover, the distilled μ-CE requires roughly 14× fewer parameters and multiply-accumulate (MAC) operations than the teacher model, supporting practical deployment for robust CSI acquisition in resource-constrained IoT and 6G edge receivers.