Method and system for adversarial audio generation for white-box scenarios

Abstract

The present invention relates to the technical field of adversarial sample generation, and in particular to a method and system for adversarial audio generation for white-box scenarios. The method can efficiently generate high-quality adversarial audio, including: selecting a target attack model and source audio and setting an attack Target; preprocess the source audio; extract the MFCC features of the source audio; the target attack model identifies the source audio according to the MFCC features, obtains the recognition result, calculates the CTC loss function between the recognition result and the attack target, and uses the particle swarm optimization algorithm Optimization, looking for the best noise, adding the best noise to the source audio to get the intermediate audio and using the target attack model to identify; if the recognition result is the same as the attack target, the intermediate audio is the confrontation audio; if the recognition result is different from the attack target, then Execute the next step; use the gradient descent algorithm to find the best noise in the middle audio until the recognition result is the same as the attack target, fine-tune the best noise and add the middle audio to get the confrontation audio.

Description

Technical field [0001] The present invention relates to the field of antagonist production, and in particular, to a counter-antiferous generation method and system for white box scenes. Background technique [0002] With the development of machine learning and artificial intelligence, the machine learning model becomes ubiquitous and has become a core technology in many artificial intelligence devices, such as speech recognition models in a voice assistant (for example, Apple Siri, Googlenow, and Amazon Echo). Sound Event Classification Model and Video Classification of Sound Event Classification Models and Video in Intelligent Significance Locks. Although machine learning is excellent, the most recent research shows that the neural network in the machine learning model is easily fooled by an attacker, and they can force the model to generate the result of errors even produce targeted output. This attack method that is called a case-assault is proved to have a high attack success...

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