A Speech Adversarial Sample Restoration Method with High Auditory Fidelity

Abstract

The invention relates to a method for repairing speech confrontation samples with the characteristics of high auditory fidelity, and relates to the field of artificial intelligence security technology. The method includes the steps of: constructing an adversarial sample repair training data set; building an RAE network and setting network parameters; constructing a high-fidelity audio reconstruction loss, that is, a high-fidelity policy improvement based on signal mean square error; setting training parameters and training the network; using The trained RAE network repairs the adversarial samples, and judges whether the repair is successful with the help of the speech recognition model. Compared with the traditional and commonly used voice signal restoration methods at present, the audio restoration samples generated by the algorithm of the present invention have higher auditory fidelity and restoration success rate, and can be applied to countermeasure samples under lower signal-to-noise ratio conditions.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence security, in particular to a method for repairing a speech confrontation sample 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, which can induce speech recognition algorithms to produce incorrect recognition results by adding weak noises that are difficult for humans to detect in the input audio data. Facing the threat of voice countermeasures, scholars at home and abroad are actively exploring the security defense technology of artificial intelligence algorithms. The main defense methods are divided into three categories. One is to expand the training data set through adversarial samples and then conduct adversarial training to improve the robustness of the intelligent algorithm. This method needs to generate a large number of adv...

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