Efficient Power Adaptation against Deep Learning Based Predictive Adversaries

Abstract

Wireless communication networks are subject to various types of adversarial attacks, which might be passive in the form of eavesdropping, or active in the form of jamming. For the former category, even if the traffic is encrypted, an adversary performing analysis on observed traffic signatures may lead to leakage of the so called contextual information regarding the traffic. New advances in the field of machine learning also result in significantly more complex adversarial units, which may deduce different forms and uses of such contextual information. In this work, we are interested in power adaptation against an intelligent adversary which utilizes deep learning and attempts to perform predictions and time forecasting on the observed traffic traces to estimate the imminent traffic intensities. Based on its traffic predictions, the adversary might possibly activate its jamming mode and utilize its limited power more efficiently to inflict maximal damage. As a method of mitigation, the transmitter may want to increase transmitter power if it expects a higher probability of jamming, and it has a significant amount of upcoming data to transmit. We leverage Lyapunov optimization and virtual queues to meet a certain level of data transmission reliability while also minimizing power consumption.