A Countermeasure Against Adversarial Attacks on Power Allocation in a Massive MIMO Network

Abstract

Deep learning has been emerging as a powerful design tool for the current and future generations of wireless networks. Among many other successful applications, deep learning has been shown to reduce computational complexity in power allocation problems in massive multiple-input and multiple-output (MIMO) networks. Despite its advantages over conventional power allocations, a recent study demonstrated that an imperceptible yet carefully designed feature perturbation named as adversarial examples may drastically degrade the performance of the power allocation system based on deep learning. Hence, in this paper, a defence system called noise-augmented neural network is investigated to mitigate the effect of adversarial attacks, and its performance against white-box fast gradient sign attacks and projected gradient descent attacks is evaluated. It is shown that the proposed noise-augmented neural network could protect power allocation system from the damaging effect of the adversarial perturbations with much greater accuracy as compared to the undefended deep neural network.