One-Shot Modulation Recognition Using Artificial Modulation Types for Enhanced Meta-Learning

Automatic modulation recognition (AMR) is of great importance for military signal reconnaissance. Here, we consider one-shot AMR, which aims at recognizing the modulation type of intercepted signals based solely on one previous recording of such signals, in terms of a, often small, number of I/Q-samples. In this paper, we introduce the novel idea of training meta-learning algorithms for one-shot AMR on large numbers of artificially generated modulation types, thereby improving the one-shot generalization performance. In comparison, previous research within the area of meta-learning AMR rely mainly on traditional AMR datasets, containing too few modulation types for optimal meta-training. Instead, we generate 675 artificial modulation types, divided into disjoint training, validation and testing sets, thus enabling a straight forward training procedure where we monitor the one-shot performance on validation classes during the training and later evaluate said performance on yet another set of unseen classes. By training a relation network on the artificial modulation types, we demonstrate a good generalization performance to both real and artificial modulation types, and can empirically show that having more classes in the training set results in a higher average five-way one-shot prediction accuracy on unseen modulation types during testing, demonstrating the need for a large number of classes during meta-training.