A Survey of Deep Transfer Learning in Automatic Modulation Classification

Abstract

Automatic modulation classification (AMC) is pivotal in both cooperative and non-cooperative communication systems. Despite achieving significant success in this field, deep learning (DL) is challenging to adapt to varying modulation parameters and channel conditions for its reliance on training data. Deep transfer learning (DTL) emerges as a potent tool to address the distribution divergence of the training and testing data, demonstrated by its successful applications in computer vision (CV). Since research on DTL within signal processing remains infantile, this paper offers a comprehensive review of current state-of-the-art (SOTA) research on DTL in modulation classification. The background and theoretical models of DTL are firstly illustrated. Specially, a detailed analysis of how transmission and reception impact data probability density function (PDF) is demonstrated for the first time. Through analyzing current literature, we identify three key classification criteria for DTL-based AMC: 1) what to transfer, 2) relationship between source and target domains or tasks, and 3) availability of labels. In addressing what to transfer, we provide a detailed conclusion and comparison of subclasses, including model-parameter-transfer, feature-representation-transfer, and instance-transfer methods. Lastly, this review discusses recent research trends and outlines future directions for DTL in AMC.