GAN-based speech confrontation sample generation method

Abstract

The invention relates to a GAN-based speech confrontation sample generation method, which is characterized by comprising the steps of preprocessing an original speech data sample x, inputting the preprocessed original speech data sample x into a generator G to obtain an adversarial disturbance G(x), and using a formula (1) to construct an adversarial sample, the formula (1) being xadv = x + G(x),inputting the adversarial sample xadv into a discriminator D, and inputting the adversarial sample xadv into a target network f after the adversarial sample xadv passes through a Mel-frequency cepstrum coefficient MFCC feature extractor, calculating the loss lf of the target network, the adversarial loss lGAN of the discriminator, the hinge loss lhinge, the mean square error loss l2 and the loss lD of the discriminator, thereby obtaining a loss function l when the generator G is trained, S4, updating parameters of a generator and a discriminator through gradient back propagation of the loss function l obtained in the S4, obtaining an optimal generator through a formula (10), loading an original sample x into the optimal generator obtained in the S5 through the formula (10), and constructing to obtain a corresponding adversarial sample. Thus, the minimum disturbance can be effectively generated, and the speech quality can be ensured.

Description

technical field [0001] The present invention relates to the field of speech technology, in particular to a GAN-based speech adversarial sample generation method. Background technique [0002] An adversarial example refers to a sample that is purposefully added with subtle perturbations. Its main purpose is to cause the performance of the deep neural network to fail, and even induce the deep learning network to make the judgment specified by the attacker. The method of constructing adversarial samples is actually a process of seeking the optimal perturbation. At present, the more common adversarial sample generation methods are divided into methods based on optimization against perturbation and methods based on perturbation. [0003] Using an optimization algorithm to find an adversarial perturbation is usually to set a target optimization function that satisfies the adversarial sample conditions, and find the optimal perturbation that satisfies the constraints; the method ba...

AI Technical Summary

Problems solved by technology

[0004] However, Alantot only considers that the adversarial samples can successfully attack the target network, and does not guarantee the quality of the generated speech; the method adopted by Carlini must first convert the speech into MFCC, then modify the MFCC features through the gradient information returned by the speech recognition network, and finally convert the MFCC features Reconstructed into a speech signal, although this can make the target speech recognition network recognize errors, but it cannot guarantee the speech quality of the confrontation sample

Images