Breaking the Performance Gap of Fully and Semisupervised Learning in Electromagnetic Signature Recognition

Abstract

Intelligent electromagnetic signature recognition is one of the key technologies in Internet of Things (IoT) device connection, which can improve system security and speed up the authentication process. In practical scenarios, as the number of IoT devices increases, electromagnetic features, such as fingerprint and modulation signals also increase substantially. However, since intelligent recognition technology, such as automatic modulation classification (AMC), requires a large amount of labeled data to train the neural network classifier, it is challenging to collect so much labeled data. To address the performance degradation challenges with small training data, we propose an efficient semisupervised electromagnetic recognition framework to break the performance gap with the fully supervised learning scheme. This framework can fully use the unlabeled electromagnetic data collected during the authentication process for self-training to improve the classifier's performance. According to the idea of consistency regularization, we design a signal augmentation method and propose an ensemble pseudolabel design algorithm to improve confidence. Moreover, we perform a convex combination of electromagnetic features to smooth the model decision boundary while generalizing to unknown data distribution regions. Experimental results on the modulated data demonstrate the performance superiority of the proposed algorithm, i.e., use less than 5% of data with no more than 10% performance drop.