Overview of Machine Learning in Communication Systems

The advancement of artificial intelligence and machine learning (ML) have made a significant impact in many research areas during the last decade. ML techniques will most certainly be implemented in some parts of future wireless communication systems. The amount of published research about ML for communications has increased enormously during the last five years. ML provides new opportunities but also new vulnerabilities towards, for example, new types of adapted jamming and spoofing attacks and difficulties in predicting and guaranteeing system performance. This report provides an overview of existing research results about using ML in communication systems. An extensive literature review has been conducted, covering many areas of communications. The large amount of research literature demonstrates that there are various applications where ML can be exploited. However, it is important to keep in mind that this does not necessarily imply that ML is the best choice, or even that it provides any benefits compared to traditional methods, for all types of applications. Applications where ML may be beneficial are, for example, such that require reduced computational complexity with near-optimal performance or model-free learning that can be adapted to phenomena that are completely or partially unknown. Examples of such algorithms include resource allocation at different levels of complex communication networks, traffic modelling, signal detection or classification of unknown signals, and end-to-end learning of channels that are unknown or difficult to model. A crucial aspect for the development and use of ML algorithms, in any application, is the access to a sufficiently large set of training data with good enough quality. The creation of such data sets may be burdensome and costly, and it is of uttermost importance to pass that hurdle to be able to use ML algorithms in communication systems. This report has only briefly touched upon the vulnerabilities of using ML techniques for communication applications. Future studies need to put more emphasis on exploitable vulnerabilities and the consequences of these on communication performance. For example, the overall knowledge is limited of adversarial attacks on this type of algorithms and defense mechanisms against such attacks. Adversarial ML attacks may also be exploited for purposes that are benefitial in defense and security applications, such as LPI/LPD communications. Such applications should be further studied