Black box voice confrontation sample generation method with auditory concealment

Abstract

The invention relates to a black box voice confrontation sample generation method with auditory concealment, and relates to the technical field of artificial intelligence safety. According to the main technical scheme, the method comprises the steps of initializing simulated annealing parameters; reading in an original audio, and initializing an audio confrontation sample; calculating black box noise according to an input audio, and performing concealment processing, namely a time-varying noise strategy based on signal variance and concealment improvement based on an auditory effect of human ears; synthesizing a new confrontationsample by using the black box noise; and inputting a black box voice recognition model, judging whether the attack is successful or not, if the attack is successful, stopping iteration, outputting an audio confrontation sample, and if the attack is not successful, generating a new solution as an input audio according to a Markov criterion to continue iteration until the iteration is completed or the attack is successful. According to the invention, the audio confrontation sample generated by the method has high similarity with the original audio, is more in line with the auditory effect of human ears, has high concealment, and can be successfully attacked without being perceived.

Description

Technical field [0001] The present invention relates to the field of artificial intelligence safety, and more particularly to a black box speech to the anti-sample generation method having audible concealed. Background technique [0002] In recent years, with the rapid development of artificial intelligence, deep learning has been gradually applied to various fields of society, especially in many areas such as security, finance, logistics, and many commercial applications. Modern speech recognition technology has also seen a major breakthrough during this period. Due to depth learning nonlinear traits and its deep network structure, the performance of the decoder, acoustic modeling and speech information is particularly prominent. In 2018, UNIFEF proposed a depth full sequence consolidation neural network (DFCNN), using a large number of convolutions to model the entire speech signal, learn from the network configuration of image identification, each convolution layer uses a smal...

AI Technical Summary

Problems solved by technology

However, in the generation method of speech confrontation samples, if the initialization temperature is set too large or too small, it is easy to fall into local minimum

Images