Semi-Supervised Radio Signal Identification

Radio signal recognition in dense and complex multi-user spectrum environments is an important tool for optimizing spectrum utilization, identifying and minimizing interference, enforcing spectrum policy, and implementing effective radio sensing and coordination systems. Classical approaches to the problem focus on energy detection and the use of expert features and decision criteria to identify and categorize specific modulation types [2] [1]. These approaches rely on prior knowledge of signal properties, features, and decision statistics to separate known modulations and are typically derived under simplified analytic hardware, propagation, radio environment models. We recently demonstrated the viability of naive feature learning for supervised radio classification systems [14] which allows for joint feature and classifier learning given labeled datasets and examples. In this case we were able to outperform traditional expert decision statistic based classification in sensitivity and accuracy by a significant margin. This was a powerful result, providing significant performance improvements against current day solutions, but it still relied entirely on supervised learning and well curated training data. In the real world, and especially in the radio domain, we are faced with vast amounts of unlabeled example data available to our sensor and incomplete knowledge of class labels comprising ground truth. To address this problem we investigate alternative strategies for radio identification learning which rely less heavily on labeled training data and are capable of making sense of radio signals with either no or less labeled examples, potentially drastically reducing the burden of data curation on such a machine learning system for developers and maintainers, and allowing systems to recognize new signals and scale to to understand new environments over time.