Machine learning for physical-layer security: Attacks and SLP Countermeasures for Multiantenna Downlink Systems

Most physical-layer security (PLS) work employ
information theoretic metrics for performance analysis. In this
paper, however, we investigate PLS from a signal processing point
of view, where we rely on bit-error rate (BER) at the eavesdropper
(Eve) as a metric for information leakage. Meanwhile,
recently, symbol-level precoding (SLP) has been shown to provide
PLS gains as a new way for security. However, in this work, we
introduce a machine learning (ML) based attack, where we show
that even SLP schemes can be vulnerable to such attacks. Namely,
this attack manifests when an eavesdropper (Eve) utilizes ML in
order to learn the precoding pattern when precoded pilots are
sent. With this ability, an Eve can decode data with favorable
accuracy. As a countermeasure to this attack, we propose a novel
precoding design. The proposed countermeasure yields high BER
at the Eve, which makes symbol detection practically infeasible
for the latter, thus providing physical-layer security between the
base station (BS) and the users. In the numerical results, we
validate both the attack and the countermeasure, and show that
this gain in security can be achieved at the expense of only a
small additional power consumption at the transmitter.