Adversarial sample generation method based on channel perception and frequency domain conversion

Abstract

The present application belongs to the field of wireless communication and cross technology of adversarial machine learning, and relates to a kind of generation method of adversarial sample based on channel perception and frequency domain conversion, comprising: based on the input original signal data set, obtain initial adversarial sample;Iterative correction is carried out to initial adversarial sample, and in each round iteration, multiple random spectrum transformations are carried out to initial adversarial sample, and multiple time domain gradients are obtained;According to the multiple time domain gradients obtained by calculation, the average gradient of the current round is obtained;Each random spectrum transformation includes: adding Gaussian noise to the current adversarial sample and obtaining the corresponding frequency domain signal sample;The time domain gradient is calculated by perturbing the frequency domain signal sample;Based on the average gradient and the real-time perceived channel matrix, the current adversarial sample is updated, and after iterative correction for multiple rounds, the optimal adversarial sample is generated.The present application improves the attack effect and practicability of adversarial sample, and provides an important reference for the security research and defense mechanism design of wireless communication system.

Images