Signal Classifiers Using Self-organizing Maps: Performance and Robustness

This paper explores the use of self-organizing maps as a mechanism for performing unsupervised learning for signal classification. Approaches using unsupervised learning have a key advantage over traditional approaches that utilize neural networks and support vector machines because they do not require a training phase. We develop signal classifiers using self-organizing maps and explore their robustness. Another concern with using unsupervised learning is the ability for an adversary to shape what is learned. In this paper we explore avenues for attack, and how they affect the performance of the signal classifier. This paper extends previous work on this topic by building a full signal classifier and quantitatively measuring its performance, and introducing two new types of attacks against classification engines.