Blind Physical Layer Authentication over Fading Wireless Channels through Machine Learning

The problem of determining the source of a message in a wireless communication link is challenging, especially for those systems in which cryptographic approaches are barely feasible due to limited resources. In this paper we consider a physical layer authentication protocol based on the characteristics of the communication channel and exploiting machine learning techniques to obtain authentication without needing any statistical knowledge of the channel from the authenticator. Different operational conditions are taken into account, considering a set of parallel channels affected by time-varying fading and assuming correlation between an opponent's channel and the authenticator's channel. Nearest Neighbor (NN) classification is used for authentication, and since the authenticator has no access to forged messages during the training phase, one-class NN classification algorithms are considered. We show that a good secrecy performance with a small training set may be achieved, allowing detection of an attacker with a very high probability in most of the cases. On the other hand, aiming at guaranteeing security even in the case of rapidly varying channels, these techniques prove to be quite conservative, and exhibit a high probability of refusing uncertain messages even when they come from the legitimate transmitter.