AutoPilot: Steering OFDM Pilots Away From Jamming Using Deep Learning

Pilot jamming attacks are among the most cost-effective, yet successful, attacks against wireless networks designed around fixed pilot locations, e.g., IEEE 802.11a/g/p. In this paper, we aim at improving the security and robustness of those systems against this class of attacks. Specifically, we present AutoPilot, a data-driven and reconfigurable transceiver designed to combat narrowband jamming attacks, with specific focus on pilot jamming. AutoPilot embeds a customized RF front-end to effectively demodulate waveforms transmitted using non-standard compliant pilot configurations as well as two data-driven logic units that are designed to (i) adapt via pseudo-randomization pilot configurations according to the current channel conditions and attack profiles; and (ii) detect rapidly and with high accuracy pilots non conforming to standard-defined pilot sequences. We introduce and describe AutoPilot design and eventually present a prototype that is evaluated via over-the-air experiments on an actual wireless testbed. Our results show that AutoPilot can detect pilots with an accuracy as high as 99% on over-the-air unseen data and can improve network throughput under jamming attacks by 4x at high jammer powers. Finally, by randomizing pilots location, AutoPilot prevents eavesdroppers from demodulating eavesdropped packets, thus providing an increased layer of security to the network.