Adaptive Whitening and Feature Gradient Smoothing-Based Anti-Sample Attack Method for Modulated Signals in Frequency-Hopping Communication

Abstract

In modern warfare, frequency-hopping communication serves as the primary method for battlefield information transmission, with its significance continuously growing. Fighting for the control of electromagnetic power on the battlefield has become an important factor affecting the outcome of war. As communication electronic warfare evolves, jammers employing deep neural networks (DNNs) to decode frequency-hopping communication parameters for smart jamming pose a significant threat to communicators. This paper proposes a method to generate adversarial samples of frequency-hopping communication signals using adaptive whitening and feature gradient smoothing. This method targets the DNN cognitive link of the jammer, aiming to reduce modulation recognition accuracy and counteract smart interference. First, the frequency-hopping signal is adaptively whitened. Subsequently, rich spatiotemporal features are extracted from the hidden layer after inputting the signal into the deep neural network model for gradient calculation. The signal’s average feature gradient replaces the single-point gradient for iteration, enhancing anti-disturbance capabilities. Simulation results show that, compared with the existing gradient symbol attack algorithm, the attack success rate and migration rate of the adversarial samples generated by this method are greatly improved in both white box and black box scenarios.