Speech adversarial sample generation method

Abstract

The invention discloses a speech adversarial sample generation method, which comprises the steps of reading input speech data, performing preprocessing operation on the input speech data, and extracting a speech feature value of the input speech data; loading a deep neural network model and parameters of the DeepSpeech speech recognition system; meanwhile, the extracted speech characteristic valueis input into a DeepSpeech speech system; calculating probability distribution of a recognition result of each frame of input speech data, initializing an error value and an error threshold value according to CTC Loss between the result and a given target value, and performing clamping operation on the error value and the generated adversarial sample; constructing a loss function of a speech adversarial sample generation algorithm, iterating for multiple times, and updating an error value; and if the identification result of the generated adversarial sample is a given target value, reducing an error threshold to update a threshold error, continuing iteration until iteration is finished, and outputting a result. The similarity between the adversarial sample generated by the algorithm and the original sample is higher.

Description

technical field [0001] The invention belongs to an algorithm for generating an adversarial sample in the field of deep learning security, in particular to a method for generating a voice adversarial sample. Background technique [0002] In recent years, with the development of deep neural networks, deep learning has been gradually applied to various fields, especially in computer vision, speech recognition, natural language processing, etc., which have reached or even surpassed human capabilities. At the same time, the security issues brought about by deep learning are getting more and more attention. Among them, the generation method of adversarial examples has gradually become a hot issue in the field of deep learning technology security. Adversarial samples refer to the samples that the deep neural network model will make wrong judgments after adding small perturbations that are difficult to distinguish by human senses on the original data that the deep neural network mo...

AI Technical Summary

Problems solved by technology

[0010] The present invention studies the DeepSpeech speech recognition system based on the cyclic neural network model structure, and proposes a targeted speech adversarial samp...

Images