Adversarial sample attack defense method and device based on speech enhancement algorithm

Abstract

The embodiment of the invention provides an adversarial sample attack defense method and device based on a voice enhancement algorithm. The method comprises the steps of obtaining a to-be-identified voice sample and spectrum features of the to-be-identified voice sample; according to the spectrum characteristics of the to-be-identified voice sample, calculating a noise spectrum of the to-be-identified voice sample through a preset algorithm, and denoising the to-be-identified voice sample by using the estimated noise spectrum obtained through calculation to obtain a denoised voice sample, wherein the algorithm comprises a spectral subtraction method based on continuous minimum tracking and a logarithm minimum mean square error algorithm MMSE algorithm combined with speech existence probability; and recognizing the denoised voice sample through a pre-trained voice recognition model to obtain a recognition result. Therefore, after the to-be-identified voice sample is obtained and de-noising processing is carried out on the to-be-identified voice sample, the de-noised voice sample is identified, so that the voice identification accuracy is improved and the efficiency of defending against the attack of the countermeasure sample is improved.

Description

technical field [0001] The invention relates to the field of information technology, in particular to a defense method and device for adversarial sample attacks based on speech enhancement algorithms. Background technique [0002] At present, with the rapid development of speech recognition technology, its use has become more and more extensive. Speech recognition technology can provide various services for people's life, and speech recognition technology also greatly improves the efficiency of human-computer interaction. However, when an attacker generates an adversarial sample by adding some deliberately small noise to the speech sample, it may cause the carefully crafted neural network to produce wrong recognition results. [0003] However, when the existing technology recognizes speech samples, it generally recognizes the speech samples directly through the trained network model, and the recognition effect on the adversarial samples is often not good, and even speech di...

AI Technical Summary

Problems solved by technology

However, when an attacker generates an adversarial sample by adding some deliberately small noises to the speech sample, it may cause the carefully crafted neural network to produce wrong recognition results.

[0003] However, when the existing technology recognizes speech samples, it generally recognizes the speech samples directly through the trained network model. The recognition effect of the adversarial samples is often not good, and even speech distortion and recognition results may be wrong.

Images