Music embedding attack defense method for voice recognition system

Abstract

A music embedding attack defense method for a voice recognition system comprises the following steps: (1) preparing a data set and preprocessing; (2) training a voice recognition model; (3) pre-training a generative adversarial network; (4) retraining the generative adversarial network; (5) generating an adversarial sample; and (6) performing adversarial training. According to the invention, musicaudios are generated through the generative adversarial network; a transcription result of the adversarial audio is obtained by using the voice recognition model, and a target function is set by using the loss between the transcription result and the target phrase and the discrimination result of a discriminator to optimize a generator, so that the generated adversarial audio cannot be recognizedby human ears and is transcribed into the target phrase; and the capability of the speech recognition system in defending against sample attack is improved through adversarial training.

Description

technical field [0001] The invention relates to a GAN-based attack defense method for a music-embedded speech recognition system. Background technique [0002] At present, deep learning has been widely used in image recognition, speech recognition, data generation and other fields. The generative adversarial network is one of the commonly used models in the field of deep learning. Through the mutual game learning of the generative model and the discriminant model, better results can be output. [0003] The speech recognition system based on deep learning has brought great convenience to people's daily production and life, but it is also vulnerable to hidden attacks. By adding disturbances that cannot be detected by the human ear to the original audio, the audio transcription results can occur If the attacker carefully designs the perturbation to make the generated adversarial samples transcribe into the target phrase, it will very likely damage the user's personal privacy a...

AI Technical Summary

Problems solved by technology

Commonly used speech recognition defense methods include confrontation training. By adding confrontation samples to the training data set and retraining the model, the model has the ability to defend against confrontation samples. Since the above attack method is to add interference to specific audio, a large amount of audio processing needs to be generated. Adversarial examples require a large amount of calculations

Images