Practical implementation of modulation classification and adversarial attacks using Universal Software Radio Peripheral with deep learning

Deep learning (DL) has proven to be highly effective in solving classification problems, making it an ideal tool for identifying unknown modulation signals. This research aims to accurately classify signal modulation classes in over-the-air approaches using software-defined radios (SDR). The study examines a wireless communication system with a transmitter (Tx) and receiver (Rx). The process of Modulation Classification at the Rx is achieved by utilizing deep learning models such as the Convolutional Neural Network (CNN), Residual Network (ResNet), and a combination of CNN and Long Short-Term Memory (CLSTM). These models classify the modulated signals, like BPSK, QPSK, 8PSK, and 16QAM. To introduce adversarial attacks, an adversary (Ad) is added to the communication system. In this report, the Fast-Gradient Sign method (FGSM) adversarial attack method is employed to create fake signals that deceive DL-based classifiers and lead to significant reductions in the accuracies of three DL models. To combat adversarial attacks, a generative-adversarial network (GAN)-based defense is proposed to enhance the accuracy of the DL models.