Towards scalable zero-shot modulation recognition

With the advent of Dynamic Spectrum Access networks, practical modulation recognition (ModRec) has become an important problem with critical applications to spectrum enforcement and resource allocation. Existing ModRec frameworks hinge on exhaustive classifier training and require labeled observations for all modulations before recognition can be performed. As a result, a general ModRec system based on fully supervised classification quickly becomes intractable due to the emergence of new proprietary waveforms and the need to collect labeled training data for them and retrain the system. This underpins the need to perform modulation classification of previously unobserved modulations. Thus, the question emerges can learned classifiers be transferred to "unobserved" modulations? In this paper, we investigate the utility of zero-shot learning to address the above challenge in the context of ModRec. We design ModRec-0, which employs unseen modulations' constellation diagrams within a zero-shot framework to transfer learned classifiers from observed modulations. We propose several important properties to enable this transfer, such as the number of constellation points, phase and amplitude levels. Our system is flexible in that additional such properties as well as a variety of classification features can be incorporated. We evaluate our framework on synthetic and real over-the-air traces and investigate the necessary conditions such as SNR levels and observed training classes for successful transfer. We demonstrate that ModRec-0 achieves over 85% average accuracy across multiple modulation classes both in synthetic and real-world traces.