The Limits of Semi-Supervised Learning for Modulation Classification in Software-Defined Radios

Abstract

Automatic modulation classification (AMC) from raw I/Q samples is an enabling function of reconfigurable software-defined radio (SDR) receivers, spectrum monitoring, and cognitive radio: before a multi-standard platform can adapt its processing chain to a signal, it must identify what it is receiving. Labeling radio captures, however, requires expert annotation or cooperative transmitters, while unlabeled recordings accumulate nearly for free on any SDR platform. Semi-supervised learning (SSL) promises to exploit this asymmetry, and prior SSL-AMC studies report consistent gains, typically, however, against unaugmented supervised baselines, on inconsistent data splits, and with single-seed results. We re-examine the question under a rigorous protocol: one shared, SNR-stratified split for every method, a strong supervised reference using the full augmentation and regularization stack, and three seeds behind every headline number. We build a MixMatch-based recipe for raw I/Q signals (signal-domain strong augmentation, a per-(class, SNR) adaptive confidence threshold, and a Mean-Teacher consistency loss) and sweep the labeled budget on RML2016.10a. The recipe delivers large gains in the genuinely label-scarce regime: +7.8 accuracy points over the strong supervised baseline at 1% labels (200 per class), closing 44% of the gap to the fully supervised ceiling, and matching supervised training that uses five times as many labels. The advantage decays monotonically and vanishes between 5% and 10% labels, where every classical SSL baseline we evaluate (self-training, co-training, semi-supervised GAN) falls below the strong supervised reference. Ablations reveal that the gain is an interaction effect: neither strong augmentation nor Mean-Teacher helps alone; the augmentations act as the perturbation model that consistency training requires. We release code, per-SNR results, and the full protocol.