Implementation of Real-Time Adversarial Attacks on DNN-based Modulation Classifier

In this paper, we provide a hardware implementation for over-the-air (OTA) adversarial attack on a deep neural network (DNN)-based modulation classifiers. Although Automatic modulation classification (AMC) using the DNN-based method outperforms the traditional classification, it has been proven that the machine learning (ML) approaches lack robustness against adversarial attacks. Therefore, the adversarial attacks cause the loss of accuracy for the DNN-based AMC by injecting a well-designed perturbation to the wireless channels. The case study presented evaluates the adversarial attack performance and its effects on the accuracy of the DNN-classifier OTA using a universal software radio peripheral (USRP) B210. Firstly, we develop an intelligent AMC system using USRPs to classify four digitally modulated signals, namely, BPSK, QPSK, 8PSK, and 16QAM, in real-time. We consider a wireless communication system that consists of three software-defined radios (SDRs), namely, transmitter, receiver, and adversarial attack. While the Rx classifies the received signal, using a DNN-based classifier, the adversarial attack node intends to misclassify the DNN-based classifier by perturbing the input data of with an adversarial example. The developed adversarial node implements the Fast-Gradient Sign method (FGSM) to generate the needed perturbation. The results of the conducted experiment show that the DNN-based classifier achieves 97% accuracy in the absence of an adversarial node. However, after deploying the adversarial attack the classifier accuracy drops to 42%.

I. Introduction

Automatic modulation classification (AMC) plays a key role in many military and civilian communication systems [1] –[3]. In other words, in AMC systems the modulation classes of the received signals can be automatically recognized without any prior knowledge. Thus, AMC increases transmission efficiency by reducing the supplementary information needed to reconfigure its parameters. There are two classes of AMC approaches [4]. the first class is the maximum likelihood-based technique that depends on the probability density function of the received samples, and their performance depends on the considered system model’s accuracy, which makes them susceptible to the model mismatch [5], [6]. The second category deploys the feature-based learning approach that collects the features from the received data samples and uses a classifier at the receiver to determine the modulation classes [7]. Particularly, the authors have used convolutional neural networks (DNNs) for the classical task of AMC.