Voice adversarial sample repairing method with characteristic of auditory high fidelity

Abstract

The invention relates to a voice adversarial sample repairing method with the characteristic of high auditory fidelity, and relates to the technical field of artificial intelligence safety. The method comprises the steps: constructing an adversarial sample repair training data set; building an RAE network and setting network parameters; constructing high-fidelity audio reconstruction loss, namely, improving a high-fidelity strategy based on a signal mean square error; setting training parameters and training the network; and performing adversarial sample repair by using the trained RAE network, and judging whether the repair is successful or not by means of a voice recognition model. Compared with a traditional common voice signal repairing method, the audio repairing sample generated by the algorithm has high auditory fidelity and repairing success rate, and can be applied to an adversarial sample under the condition of lower signal-to-noise ratio.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence security, in particular to a method for repairing speech confrontation samples with the characteristics of high auditory fidelity. Background technique [0002] In recent years, speech recognition adversarial attacks have become a new research hotspot in artificial intelligence. It can induce speech recognition algorithms to produce false recognition results by adding weak noises that are difficult for humans to detect in the input audio data. Facing the threat of voice counterattacks, scholars at home and abroad are actively exploring artificial intelligence algorithm security defense technologies. The main defense methods are divided into three categories. One is to expand the training data set by adversarial samples and then perform adversarial training to improve the robustness of the intelligent algorithm. This method needs to generate a large number of adversarial samples and...

AI Technical Summary

Problems solved by technology

However, the shortcomings of this method are: a large number of adversarial samples need to be generated, and only one type of adversarial samples is used for adversarial training, which is not effective in defending against other attack methods; the original speech recognition algorithm needs to be retrained , which has limitations in practical applications

However, the shortcomings of this method are: the method is essentially a traditional general-purpose noise reduction algorithm, and there is no targeted improvement for adversarial sample noise, and the defense success rate is not high

Images