Defense Against Machine Learning Based Attacks in Multi-UAV Networks: A Network Coding Based Approach

Abstract

Thanks to the agility and mobility features, unmanned aerial vehicles (UAVs) have been applied for a wide range of civil and military missions. To remotely control and monitor UAVs, mission-related data such as location and trajectory information are transmitted over wireless channels. However, UAV networks are vulnerable to eavesdropping attacks due to: 1) the broadcasting nature of wireless channels; 2) the broad coverage in aerial environments. In this paper, we investigate the potential security threats in UAV networks with passive attackers who aim to eavesdrop and decode encrypted locations by using machine learning techniques. We show that a neural network of two hidden layers is able to decode the encrypted locations if using the existing location protection methods. To defend against such machine learning based attacks, we suggest a location protection approach based on the random linear network coding with encryption keys being randomly permuted. We prove that our proposed approach allows for a low attacker's success probability and provides untraceability property. Our simulation results indicate that our approach significantly outperforms the existing location protection methods in terms of attacker's bit error rate, even with a small number of UAVs.