Defense method and defense device for black box attack model of speech recognition system

Abstract

The invention discloses a defense method and a defense device for a black box attack model of a speech recognition system. According to the defense method, simulated environment noise is added to an original audio; a voice input condition is simulated in a real scene; noise is randomly added to form a primary adversarial sample; the adversarial sample is optimized through a genetic algorithm andgradient estimation to obtain an accurate adversarial sample, and then the original audio file and the adversarial sample are mixed to serve as a training data set of adversarial training to train themodel, so that the recognition accuracy of the model to the adversarial sample is improved, and the robustness of the model to the adversarial attack is improved.

Description

technical field [0001] The invention belongs to the technical field of deep learning security, and in particular relates to a defense method and a defense device for a black-box attack model of a speech recognition system. Background technique [0002] With the development of technology, the goal of modern speech recognition technology is to teach the machine to accurately recognize the speech it hears and execute instructions correctly according to the content of the speech. Speech recognition technology simplifies the communication between human and machine, omitting the steps of intermediate keyboard control and handwriting, which enables people to perform other operations with their hands or eyes while performing voice control. These advantages have made automatic speech recognition systems widely used in various fields from national defense, medical care to intelligent driving. In defense, for example, automatic speech recognition systems provide selected cockpit contr...

AI Technical Summary

Problems solved by technology

However, in real-life applications, the parameter structure of the model is usually unavailable, and the gener

Images