A A Novel LSTM-based OFDM Channel Estimation for Adversarial Attack Defending

Abstract

Abstract: In this paper, we propose a novel Long Short-Term Memory-based channel estimation
(LSTM-CE) method which can effectively detect and protect OFDM systems from adversarial attacks. Adversarial attack is a kind of pilot jamming. Generally, this attack is intentionally created
to directly attack a neural network, distorting the channel estimation and severely degrading the performance of the OFDM system. Unlike conventional neuron networks based channel estimation, in
the proposed LSTM-CE method, we define a novel loss function owning the target to optimize the
training process, and to effectively eliminate the effect of adversarial attack. In addition, the effect
of perturbations caused by adversarial attack along the time and frequency axes has been carefully
analyzed to determine the optimal LSTM model. The experimental results show that the proposed
LSTM-CE method can not only detect adversarial attacks well but also effectively exploit the relationship in time and frequency domains to improve the performance of channel estimation as compared
to other conventional methods.
Keywords: Channel Estimation, Adversarial Attack, LSTM.