Regitive audio sample generation method based on three-population parallel genetic algorithm

Abstract

The invention discloses an adversarial audio sample generation method based on a three-population parallel genetic algorithm. The adversarial audio sample generation method comprises the following steps: A, obtaining an original audio sample; b, obtaining a corresponding input sample, a main population and two auxiliary populations; c, the fitness score of each individual is calculated; d, sorting all individuals in the main population and the auxiliary population according to a sequence of fitness scores from low to high; e, using a speech recognition model to carry out classification recognition on all the sorted individuals in the main population in sequence; and F, performing intersection, variation and individual updating on the main population and the auxiliary population by using a three-population parallel genetic algorithm, and then returning to the step D. According to the method, the optimal solution meeting the requirement can be obtained through multiple iterations, the problem of unknowability of the target network and errors caused by Mel-frequency cepstrum coefficient conversion is fully solved, and the method has the advantages of being high in convergence speed, high in global search capability and high in convergence efficiency.

Description

Technical field [0001] The present invention relates to the field of speech recognition, in particular to an adversarial audio sample generation method based on three populations of parallel genetic algorithms. Background [0002] With the success of deep learning models in speech recognition applications, automatic speech recognition control systems such as Amazon Alexa, Google Voice Assistant, Apple Siri, Microsoft Cortana and iFLYTEK have been widely used in human-computer interaction, and have been successful in mobile devices, smart homes and other fields, especially in autonomous driving, voiceprint authentication and other high-level security application scenarios to achieve key applications. [0003] And recent studies have shown that neural networks are vulnerable to adversarial attacks. There is also a problem in the field of speech recognition, where an attacker adds a slight perturbation to the audio, causing the neural network to input a completely different value, ...

AI Technical Summary

Problems solved by technology

But on the one hand, compared with the generation of adversarial examples in the field of computer vision, it is more difficult to design adversarial examples in the field of speech; on the other hand, the current research on adversarial examples in the field of speech is relatively weak

At present, the adversarial attack algorithms in the voice field are all based on the optimized C&W attack algorithm design. This method often requires huge computing resources and time overhead, which seriously restricts the practicability of the current voice adversarial attack algorithms.

Images