Self-Supervised Learning and Nearest Neighbors for Out-of-Distribution Modulation Classification

Abstract

In non-cooperative communications, most modulation recognition tasks assume that the transmitted (training) signals and the received (test) signals are independent and identically distributed. However, the received signals may suffer from unknown impairments in practical communication systems, resulting in different data distributions between the test and training signals. Such test signals are regarded as out-of-distribution (OOD) signals. In this letter, we consider OOD scenarios with varying carrier frequency offset and varying sampling frequency and propose a new self-supervised method to improve the robustness of modulation recognition. The key idea is to use contrastive learning with information maximization to pre-train a feature extractor from unlabeled signals and leverage the feature space of the fine-tuned feature extractor with deep k-nearest neighbors to recognize modulation modes of the test signals. Experimental results show that our method has better representation and recognition accuracy than the baseline methods.