Confronting audio generation method and system for white-box scene

Abstract

The invention relates to the technical field of confronting sample generation, in particular to a confronting audio generation method and system for a white-box scene. The method can generate a high quality confronting audio efficiently. The method comprises the following steps: selecting a target attacking model and a source audio and setting an attacking target; pre-processing the source audio;extracting an MFCC characteristic of the source audio; recognizing the source audio by the target attacking model according to the MFCC characteristic to obtain a recognizing result, calculating a CTCloss function between the recognizing result and the attacking target and optimizing the function by means of particle swarm optimization, searching for the optimum noise, and adding the optimum noise into the source audio to obtain an intermediate audio and recognizing the noise by using the target attaching model; if the recognizing result is the same as the attacking target, the intermediate audio is the confronting audio; if the recognizing result is different from the attacking target, executing the next step; and searching for the optimum noise of the intermediate audio by using a gradient descent algorithm till the recognizing result is the same as the attacking target and finely adjusting the optimum noise and adding the intermediate audio to obtain the confronting audio.

Description

technical field [0001] The present invention relates to the technical field of adversarial sample generation, in particular to a method and system for adversarial audio generation for white-box scenarios. Background technique [0002] With the development of machine learning and artificial intelligence, machine learning models have become ubiquitous and have become the core technology in many artificial intelligence devices, such as speech recognition models in voice assistants (for example, Apple Siri, GoogleNow, and Amazon Echo), Speaker recognition models in smart voice locks, sound event classification models in acoustic surveillance systems and classification of pornographic videos. Despite machine learning's impressive performance, recent research has shown that the neural networks in machine learning models can be easily fooled by attackers, who can force the models to produce erroneous results or even targeted outputs. This attack method, called adversarial example ...

AI Technical Summary

Problems solved by technology

However, the existing white-box confrontation audio generation methods are relatively rudimentary and time-consuming, and the attack effect is poor.

Images