Investigation of Deep Learning Architectures and Features for Adversarial Machine Learning Attacks in Modulation Classifications

Artificial intelligence (AI), and specifically machine and deep learning, are emerging as essential enabling techniques for the design of future generations of wireless networks for the Internet of Things (IoT). With the exponential increase in IoT devices in recent years, AI methods have become even more beneficial in network management including for energy efficiency, spectrum utilisation and user admission control. However, the rise of AI applications, has also caused a rise in cyber-attacks, where attackers can exploit network vulnerabilities through its use. In this research, we develop a deep learning approach for Automatic Modulation Classification (AMC) with three different feature combinations, using batch normalisation and optimised with the focal loss. Our results have implications about the performance of the AMC against adversarial examples. Firstly, we find that using batch normalisation enhances the classifier's performance against adversarial examples. Secondly, by generating white-box fast gradient sign method attacks, we show that using the phase as an input feature improves the performance of the AMC against adversarial examples.