This paper analyzes the worst-case detection performance of a feature-based physical layer authentication (PLA) scheme subject to optimal multiple-antenna impersonation attacks. The PLA scheme is based on the location-specific channel in the uplink towards a multiple-antenna receiver, and the attacker is using pre-coding with the objective of maximizing the missed detection probability. We solve the optimal attack strategy problem under perfect channel-state information (CSI) at the attacker, imperfect CSI at the attacker, and for a power constrained attacker. As a counter strategy, we propose to reserve a subset of silent receive antennas for reception only, in order to limit the CSI that an attacker can extract from overhearing downlink transmissions. We evaluate the performance under the attack- and counter-strategies, both analytically and for recorded real-world channel traces, and show that the worst-case performance is determined by the feature-energy outside the attacker's channel range and the attack-power constraints. Results indicate that an unconstrained attacker with favorable conditions can achieve a success probability close to 1; however, under more realistic channel constraints, detection performance guarantees in the order of 10(-6) - 10(-4) can be obtained. Moreover, we find that performance can be improved by 1-2 orders of magnitude through the proposed counter strategy.