A method for generating adversarial audio samples based on three-population parallel genetic algorithm

Generating adversarial audio samples through three population parallel genetic algorithms solves the problem of high computing resources for adversarial sample generation in the field of speech recognition, and achieves rapid convergence and efficient adversarial audio sample generation, which enhances the resistance of the speech recognition model.

CN114373467BActive Publication Date: 2024-11-26SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210026272.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-11
Publication Date
2024-11-26
Estimated Expiration
2042-01-11

AI Technical Summary

Technical Problem

The existing adversarial sample generation method in the field of speech recognition consumes a lot of computing resources, is serious time overhead, and is difficult to achieve practicality. Compared with the difficulty in designing in the field of computer vision, the adversarial attack algorithm in the field of speech is relatively weak.

Method used

Adversarial audio sample generation method based on three population parallel genetic algorithms is adopted to obtain the optimal solution through multiple iterations, and the adversarial audio samples are generated using Gaussian noise and genetic operations, which improves the convergence speed and global search ability, and solves the Mel frequency. Error problems caused by cepspectral coefficient conversion.

Benefits of technology

It reduces the computational amount and time of adversarial sample generation, improves global search capabilities, realizes fast convergence and efficient adversarial audio sample generation, and enhances the resistance of the target network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114373467B_ABST
    Figure CN114373467B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for generating adversarial audio samples based on a three-population parallel genetic algorithm, comprising the following steps: A: obtaining an original audio sample; B: obtaining a corresponding input sample and a main population and two auxiliary populations; C: calculating the fitness score of each individual; D: sorting all individuals in the main population and the auxiliary population in order of fitness scores from low to high; E: using a speech recognition model to classify and identify all sorted individuals in the main population in turn; F: using a three-population parallel genetic algorithm to perform intersection, mutation and individual update on the main population and the auxiliary population, and then returning to step D. The present invention can obtain an optimal solution that meets the requirements through multiple iterations, fully solves the unknowability of the target network and the error problem caused by the conversion of Mel frequency cepstral coefficients, and has the advantages of fast convergence speed, strong global search capability and high convergence efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of speech recognition, and in particular to a method for generating adversarial audio samples based on a three-population parallel genetic algorithm. Background Art

[0002] With the success of deep learning models in speech recognition applications, automatic speech recognition control systems, such as commercial products such as Amazon Alexa, Google Voice Assistant, Apple Siri, Microsoft Cortana and iFlytek, have been widely used in human-computer interaction and have achieved success in many fields such as mobile devices and smart homes, especially in application scenarios with higher security levels such as autonomous driving and voiceprint authentication.

[0003] Recent studies have shown that neural networks are vulnerable to adversarial attacks. This problem also exists in the field of speech recognition. When an attacker adds slight disturbances to the audio, the neural network inputs completely different values, but the human ear cannot recognize slight disturbances. Training the model by generating adversarial samples can effectively improve the ability of neural networks to resist malicious attacks. However, on the one hand, compared with the generation of adversarial samples in the field of computer vision, it is more difficult to design adversarial samples in the field of speech; on the other hand, current research on adversarial samples in the field of speech is relatively weak. At present, adversarial attack algorithms in the field of speech are all designed based on the optimized C&W attack algorithm. This method often requires huge computing resources and time overhead, which seriously restricts the practicality of current speech adversarial attack algorithms. Summary of the invention

[0004] The purpose of the present invention is to provide an adversarial audio sample generation method based on a three-population parallel genetic algorithm, which can obtain an optimal solution that meets the requirements through multiple iterations, fully solves the unknowability of the target network and the error problem caused by Mel-frequency cepstral coefficient (MFCC) conversion, and has the advantages of fast convergence speed, strong global search capability and high convergence efficiency.

[0005] The present invention adopts the following technical solutions:

[0006] A method for generating adversarial audio samples based on a three-population parallel genetic algorithm comprises the following steps:

[0007] A: Initialize each original speech file in the speech data set into an original audio sample in the form of a binary string; then proceed to step B;

[0008] B: Select an original audio sample, repeatedly add Gaussian noise to the least significant bit of a random subset of the original audio sample N times, and obtain corresponding N generated audio samples, i.e., input samples; according to this method, obtain a main population and two auxiliary populations for each original audio sample, each population consists of N generated audio samples; then proceed to step C;

[0009] C: Treat each input sample as an individual and calculate the fitness score of each individual; then proceed to step D; the fitness score of the individual is the Euclidean distance between the original audio sample and the generated audio sample corresponding to the individual;

[0010] D: Use the fitness score to sort all individuals in the main population and the auxiliary population in order from low to high fitness scores; then go to step E;

[0011] E: Use the speech recognition model to classify and identify all the sorted individuals in the main population in turn. During the classification and identification process:

[0012] If an individual with a successful attack appears in the main population, the classification and recognition is stopped, and the individual with a successful attack is directly output as the final adversarial audio sample;

[0013] If no individuals successfully attack in the main population, determine whether the set number of iterations has been reached:

[0014] If the number of iterations is reached, the judgment process is exited and the result of failure in generating adversarial audio is output;

[0015] If the number of iterations has not been reached, go to step F;

[0016] F: The individuals in the main population and the two auxiliary populations are divided into elite individuals and non-elite individuals according to the set retention probability. After setting different genetic operation parameters, the main population and the two auxiliary populations are subjected to genetic operations including crossover and mutation operations, wherein the elite individuals are not subjected to genetic operations; the main population and the two auxiliary populations that have completed the genetic operation are obtained, and the offspring individuals corresponding to each population are obtained, and the offspring individuals are composed of elite individuals that have not undergone crossover and mutation operations and non-elite individuals that have undergone crossover and mutation operations; then the fitness scores of each offspring individual calculated are sorted in order from low to high in the corresponding population that has completed the genetic operation; then, according to the set optimal individual selection threshold, several optimal individuals are selected from each of the two auxiliary populations that have completed the genetic operation, and added to the main population that has completed the genetic operation, and the corresponding number of offspring individuals with the highest fitness scores in the main population that has completed the genetic operation are replaced according to the fitness scores, so as to obtain the main population that has completed the genetic operation and individual update; then return to step D.

[0017] In the step A, the speech recognition model adopts tensorflow's official speech_commands. The speech recognition model is used to perform recognition training on 10 groups of speech files that have been classified by labels. The label of each group of speech files is a corresponding English word, and each group of speech files includes speech files in which the English word is read out by different speakers.

[0018] In step B, the individuals and numbers in the main population and the two auxiliary populations are the same, and the number of individuals in a single population is set to 20-40.

[0019] In the step B, when adding Gaussian noise, each bit element of the original audio sample in the form of a binary string is traversed, and the currently traversed element is converted from 1 to 0 or from 0 to 1 with a set conversion probability.

[0020] The step F comprises the following specific steps:

[0021] F1: Using the elitist method, the first P individuals of the sorted contemporary individuals in the main population and the two auxiliary populations are selected as elite individuals with a set retention probability, and the remaining NP individuals are selected as non-elite individuals;

[0022] F2: For the main population and two auxiliary populations composed of contemporary individuals, the average crossover method is used to perform crossover operations, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the crossover operation; then enter step F3;

[0023] F3: Perform mutation operation on the main population and two auxiliary populations that have completed the crossover operation, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the mutation operation; then enter step F4;

[0024] F4: Using the method in step D, first calculate the fitness score of each offspring individual in the main population and the two auxiliary populations that have completed the genetic operation; then sort all offspring individuals in the main population and the two auxiliary populations that have completed the previous round of genetic operations in order of fitness scores from low to high; then, according to the set optimal individual selection threshold, the first L individuals in each auxiliary population are sorted as the optimal individuals, and the 2L optimal individuals are added to the main population that has completed the genetic operation; then, in order of fitness scores from low to high, all offspring individuals in the main population that has been added with the 2L optimal individuals are sorted, and the 2L individuals with the highest fitness scores that are ranked last in the main population are eliminated; finally, the main population with the 2L offspring individuals eliminated is used as the main population that has completed the genetic operation and individual update; then return to D.

[0025] In the step F, the average crossover method is used to perform the crossover operation; when performing the genetic operation, a three-population parallel genetic algorithm is used; a main population and two auxiliary populations are independent of each other, different populations perform different genetic operations, and no information exchange is performed between the three populations; for the main population and the two auxiliary populations, the genetic operation is performed after setting different genetic operation parameters.

[0026] In the step F, the first auxiliary population is set to have a small mutation probability and a large crossover probability, that is, the mutation probability of the first auxiliary population is the minimum mutation probability among the three populations, and the crossover probability of the first auxiliary population is the maximum crossover probability among the three populations; the second auxiliary population is set to have a small crossover probability and a large mutation probability, that is, the crossover probability of the second auxiliary population is the minimum crossover probability among the three populations, and the mutation probability of the second auxiliary population is the maximum mutation probability among the three populations.

[0027] In the step F1, when performing a crossover operation on the population, two contemporary individuals A and B are randomly selected from the non-elite individuals in the population as parent individuals, and then each element in the contemporary individuals A and B is traversed, and the elements at the corresponding positions of the contemporary individuals A and B are cross-exchanged according to the set population crossover probability threshold to complete the crossover operation of the population and obtain the contemporary non-elite individuals generated after the population crossover.

[0028] In step F3, when performing mutation operation on the population, for each contemporary non-elite individual generated after the crossover of the population, each element of the contemporary non-elite individual is traversed, and through the set population mutation threshold, some elements in the contemporary non-elite individual are mutated to complete the population mutation operation, and obtain the population that has completed the genetic operation, and the offspring individuals corresponding to each population; the offspring individuals in the main population that has completed the genetic operation and the two auxiliary populations are composed of elite individuals that have not undergone crossover mutation operations and non-elite individuals that have undergone crossover mutation operations.

[0029] Main population crossover probability threshold M C The range is 40%≤M C ≤60%; First auxiliary population crossover probability threshold A 1C The range is 60% 1C ≤90%; Second auxiliary population crossover probability threshold A 2C The range is 10% ≤ A 2C <40%; among them, M C M is the abbreviation of main, A 1C and A 2C The A in the figure is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population, and the c in the subscripts are the abbreviations of crossover;​

[0030] Main population mutation probability threshold M m The range is 0.0004≤M m ≤0.0006; first auxiliary population mutation probability threshold A 1m The range is 0.0001≤A 1m <0.0004; the second auxiliary population mutation probability threshold A 2m The range is 0.0006 2m ≤0.0009; where M m M is the abbreviation of main, A 1m and A 2m The A in the expression is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population, and the m in the subscripts are the abbreviations of mutation.

[0031] The present invention is based on a three-population parallel genetic algorithm, and can obtain an optimal solution that meets the requirements through multiple iterations, fully solving the unknowability of the target network and the error problem caused by Mel-frequency cepstral coefficient (MFCC) conversion, and has the advantages of fast convergence speed, strong global search capability, and high convergence efficiency, reducing the amount of calculation and time for adversarial sample generation, and effectively improving the global search capability. The present invention effectively solves the problem that the traditional genetic algorithm has only one population, which makes it difficult to ensure population diversity, and the algorithm is easy to fall into the local optimum and cannot obtain the global optimal solution by means of three-population parallelism. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below with reference to the accompanying drawings and embodiments:

[0034] like Figure 1 As shown, the adversarial audio sample generation method based on three-population parallel genetic algorithm of the present invention comprises the following steps:

[0035] A: Initialize each original speech file in the speech data set into an original audio sample in the form of a binary string; then proceed to step B;

[0036] ​In the present invention, the speech recognition model uses the official speech_commands of tensorflow. Using this speech recognition model, 10 groups of speech files that have been classified by labels are trained for recognition. The label of each group of speech files is a corresponding English word, such as go or stop, etc. Each group of speech files includes speech files in which different speakers read out the English word. In this embodiment, the number of each group of speech files contains no less than 1700.

[0037] B: Select an original audio sample, repeatedly add Gaussian noise to the least significant bit of a random subset of the original audio sample N times, and obtain corresponding N generated audio samples, i.e., input samples; according to this method, obtain a main population and two auxiliary populations for each original audio sample, each population consists of N generated audio samples; then proceed to step C;

[0038] In this embodiment, an original audio sample obtained in step A is selected, and Gaussian noise is added to the least significant bit of a random subset of the original audio sample to obtain a generated audio sample corresponding to the original audio sample, that is, an input sample; then, Gaussian noise is repeatedly added to the original audio sample N times according to this method to obtain N generated audio samples corresponding to the original audio sample; and a main population and two auxiliary populations are generated using the populations composed of the N generated audio samples, respectively, the individuals and numbers in the main population and the two auxiliary populations are the same, and the number of individuals in a single population is set to 20-40; using the above method, a main population and two auxiliary populations for each original audio sample are finally obtained.

[0039] When adding Gaussian noise, each element of the original audio sample in the form of a binary string is traversed, and the currently traversed element is converted from 1 to 0 or from 0 to 1 with a set conversion probability; finally, a generated audio sample corresponding to the original audio sample is obtained. In this embodiment, the set conversion probability is 0.0001.

[0040] C: Treat each input sample as an individual and calculate the fitness score of each individual; then proceed to step D;

[0041] In the present invention, the Euclidean distance between the original audio sample and the generated audio sample is used as the fitness score of the individual. The greater the Euclidean distance between the original audio sample and the generated audio sample, the higher the fitness score of the individual, and the lower the acoustic similarity between the two.

[0042] D: Use the fitness score to sort all individuals in the main population and the auxiliary population in order from low to high fitness scores; then go to step E;

[0043] Assume that there are N individuals in the main population and each auxiliary population. The individuals with lower fitness scores have higher acoustic similarity and are ranked higher.

[0044] E: Use the speech recognition model to classify and identify all the sorted individuals in the main population in turn. During the classification and identification process:

[0045] If an individual with a successful attack appears in the main population, the classification and recognition is stopped, and the individual with a successful attack is directly output as the final adversarial audio sample;

[0046] If no individuals successfully attack in the main population, determine whether the set number of iterations has been reached:

[0047] If the number of iterations is reached, the judgment process is exited and the result of failure in generating adversarial audio is output;

[0048] If the number of iterations has not been reached, go to step F;

[0049] In the present invention, there are two attack modes: non-target attack and target attack:

[0050] For non-targeted attacks, for a certain individual, the attack is successful if the speech recognition model recognizes the generated audio sample corresponding to the individual as any other label.

[0051] For targeted attacks, for a certain individual, the speech recognition model can identify the generated audio sample corresponding to the individual as another specified label, which means the attack is successful.

[0052] F: The individuals in the main population and the two auxiliary populations are divided into elite individuals and non-elite individuals according to the set retention probability. After setting different genetic operation parameters, the main population and the two auxiliary populations are subjected to genetic operations including crossover and mutation operations, wherein the elite individuals are not subjected to genetic operations; the main population and the two auxiliary populations that have completed the genetic operation are obtained, and the offspring individuals corresponding to each population are obtained, and the offspring individuals are composed of elite individuals that have not undergone crossover and mutation operations and non-elite individuals that have undergone crossover and mutation operations; then the fitness scores of each offspring individual calculated are sorted in order from low to high in the corresponding population that has completed the genetic operation; then, according to the set optimal individual selection threshold, several optimal individuals are selected from each of the two auxiliary populations that have completed the genetic operation, and added to the main population that has completed the genetic operation, and the corresponding number of offspring individuals with the highest fitness scores in the main population that has completed the genetic operation are replaced according to the fitness scores, so as to obtain the main population that has completed the genetic operation and individual update; then return to step D.

[0053] In the present invention, the crossover operation is to use two contemporary individuals A and B in the form of binary strings as parent individuals, and exchange the elements at the corresponding positions of contemporary individuals A and B to generate offspring individuals. The crossover operation is divided into single-point crossover, multi-point crossover and average crossover according to different crossover methods. In the present invention, the average crossover method is used for crossover operation, and the average crossover is a crossover operation method in which each bit corresponding to the two parent individuals has a certain probability of being exchanged. The mutation operation is to mutate some elements in the contemporary individual in the form of a binary string after the crossover operation is performed, changing from 1 to 0, or from 0 to 1.

[0054] Since the traditional genetic algorithm has only one population, all genetic operators are only for individuals in one population, and the diversity of the population is difficult to ensure, and it is very easy to fall into the local optimum and fail to obtain the global optimal solution. Without solving the above problem, the present invention adopts a three-population parallel genetic algorithm: a main population and two auxiliary populations are independent of each other, different populations perform different genetic operations, and no information exchange is performed between the three populations. For the main population and the two auxiliary populations, genetic operations are performed after setting different genetic operation parameters.

[0055] In the present invention, the first auxiliary population focuses on global search capability and is set to have a small mutation probability and a large crossover probability, that is, the mutation probability of the first auxiliary population is the minimum mutation probability among the three populations, and the crossover probability of the first auxiliary population is the maximum crossover probability among the three populations; the second auxiliary population focuses on local search capability and is set to have a small crossover probability and a large mutation probability, that is, the crossover probability of the second auxiliary population is the minimum crossover probability among the three populations, and the mutation probability of the second auxiliary population is the maximum mutation probability among the three populations.

[0056] In the present invention, the step F comprises the following specific steps:

[0057] F1: Using the elitist method, the first P individuals in the sorted contemporary individuals in the main population and the two auxiliary populations are selected as elite individuals with a set retention probability, and the remaining NP individuals are selected as non-elite individuals, where the value range of P is 2-4 and the set retention probability is 90%.

[0058] Elitism is an optimization of the basic genetic algorithm. Since the crossover and mutation operators are performed randomly, it is possible to evolve in a good direction or a bad direction, resulting in the population losing the best individuals during the evolution process, which may lead to a decrease in the fitness function. In order to prevent the optimal solution generated during the evolution process from being destroyed by crossover and mutation, the optimal solution in each generation is copied intact to the next generation. Therefore, the elitist method is used in the present invention to retain the best elite individuals of the population with a high probability, and no operation is performed on the elite individuals, which are directly added to the offspring population.

[0059] F2: For the main population and two auxiliary populations composed of contemporary individuals, the average crossover method is used to perform crossover operations, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the crossover operation; then enter step F3;

[0060] When performing a crossover operation on a population, two contemporary individuals A and B are randomly selected from the non-elite individuals in the population as parent individuals, and then each element in the contemporary individuals A and B is traversed, and the elements at the corresponding positions of the contemporary individuals A and B are cross-exchanged according to the set population crossover probability threshold to complete the population crossover operation and obtain the contemporary non-elite individuals generated after the population crossover;

[0061] Among them, the main population crossover probability threshold M C The range is 40%≤M C ≤60%; First auxiliary population crossover probability threshold A 1C The range is 60% 1C ≤90%; Second auxiliary population crossover probability threshold A 2C The range is 10% ≤ A 2C <40%;

[0062] In this embodiment, the main population crossover probability threshold M is set C , the first auxiliary population crossover probability threshold A 1C and the second auxiliary population crossover probability threshold A 2c 50%, 70% and 30% respectively. C M is the abbreviation of main, A 1C and A 2C The A in the figure is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population, and the c in the subscripts are the abbreviations of crossover;

[0063] F3: Perform mutation operation on the main population and two auxiliary populations that have completed the crossover operation, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the mutation operation; then enter step F4;

[0064] When performing mutation operations on the population, for each contemporary non-elite individual generated after the crossover of each population, each element of the contemporary non-elite individual is traversed, and through the set population mutation threshold, some elements in the contemporary non-elite individual are mutated to complete the population mutation operation, and the population that has completed the genetic operation and the corresponding offspring individuals of each population are obtained; the offspring individuals in the main population that has completed the genetic operation and the two auxiliary populations are composed of elite individuals that have not undergone crossover mutation operations and non-elite individuals that have undergone crossover mutation operations; ​

[0065] Among them, the main population mutation probability threshold M m The range is 0.0004≤M m ≤0.0006; During the mutation process, each element of the contemporary non-elite individuals generated after population crossover has M m The mutation rate of the first auxiliary population. 1m The range is 0.0001≤A 1m <0.0004; the second auxiliary population mutation probability threshold A 2m The range is 0.0006 2m ≤0.0009. Among them, M m M is the abbreviation of main, A 1m and A 2m The A in the figure is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population, and the m in the subscripts are the abbreviations of mutation.

[0066] In this embodiment, the main population mutation probability threshold M is set m , the first auxiliary population mutation probability threshold A 1m and the second auxiliary population mutation probability threshold A 2m They are 0.0005, 0.0001 and 0.0009 respectively.

[0067] F4: Using the method in step D, first calculate the fitness score of each offspring individual in the main population and the two auxiliary populations that have completed the genetic operation; then sort all offspring individuals in the main population and the two auxiliary populations that have completed the previous round of genetic operations in order of fitness scores from low to high; then, according to the set optimal individual selection threshold, the first L individuals in each auxiliary population are sorted as the optimal individuals, and the 2L optimal individuals are added to the main population that has completed the genetic operation; then, in order of fitness scores from low to high, all offspring individuals in the main population that has been added with the 2L optimal individuals are sorted, and the 2L individuals with the highest fitness scores that are sorted last in the main population are eliminated; finally, the main population with the 2L offspring individuals eliminated is used as the main population that has completed the genetic operation and individual update; then return to D, where the value of L is 1 or 2.

[0068] After returning to step D, all individuals in the main population and auxiliary population that have completed genetic operations and individual updates are sorted in order of fitness scores from low to high using the fitness scores; then proceed to step E;

[0069] After entering step E, the speech recognition model is used to classify and identify all the sorted individuals in the main population in turn. During the classification and identification process:​

[0070] If an individual with a successful attack appears in the main population, the classification and recognition is stopped, and the individual with a successful attack is directly output as the final adversarial audio sample;

[0071] If no individuals successfully attack in the main population, determine whether the set number of iterations has been reached:

[0072] If the number of iterations is reached, the judgment process is exited and the result of failure in generating adversarial audio is output;

[0073] If the number of iterations has not been reached, go to step F;

[0074] According to the above method, steps D to F are repeatedly executed until an individual with a successful attack is found and output as the final adversarial audio sample; or when the number of iterations is reached and no individual with a successful attack is found, the judgment process is exited and the result of failure to generate adversarial audio is output.

Claims

1. A method for generating adversarial audio samples based on a three-population parallel genetic algorithm, characterized in that: The following steps are involved: A: Initialize each original speech file in the speech data set into an original audio sample in the form of a binary string; then proceed to step B; B: Select an original audio sample, repeatedly add Gaussian noise to the least significant bit of a random subset of the original audio sample N times, and obtain corresponding N generated audio samples, i.e., input samples; according to this method, obtain a main population and two auxiliary populations for each original audio sample, each population consists of N generated audio samples; then proceed to step C; C: Treat each input sample as an individual and calculate the fitness score of each individual; Then proceed to step D; the fitness score of the individual is the Euclidean distance between the original audio sample and the generated audio sample corresponding to the individual; D: Use the fitness score to sort all individuals in the main population and the auxiliary population in order from low to high fitness scores; then go to step E; E: Use the speech recognition model to classify and identify all the sorted individuals in the main population in turn. During the classification and identification process: If an individual with a successful attack appears in the main population, the classification and recognition is stopped, and the individual with a successful attack is directly output as the final adversarial audio sample; If no individuals successfully attack in the main population, determine whether the set number of iterations has been reached: If the number of iterations is reached, the judgment process is exited and the result of failure in generating adversarial audio is output; If the number of iterations has not been reached, go to step F; F: The individuals in the main population and the two auxiliary populations are divided into elite individuals and non-elite individuals according to the set retention probability. After setting different genetic operation parameters, the main population and the two auxiliary populations are subjected to genetic operations including crossover and mutation operations, wherein the elite individuals are not subjected to genetic operations; the main population and the two auxiliary populations that have completed the genetic operation are obtained, and the offspring individuals corresponding to each population are obtained, and the offspring individuals are composed of elite individuals that have not undergone crossover and mutation operations and non-elite individuals that have undergone crossover and mutation operations; then the fitness scores of each offspring individual calculated are sorted in order from low to high in the corresponding population that has completed the genetic operation; then, according to the set optimal individual selection threshold, several optimal individuals are selected from each of the two auxiliary populations that have completed the genetic operation, and added to the main population that has completed the genetic operation, and the corresponding number of offspring individuals with the highest fitness scores in the main population that has completed the genetic operation are replaced according to the fitness scores, so as to obtain the main population that has completed the genetic operation and individual update; then return to step D.

2. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 1, characterized in that: In the step A, the speech recognition model adopts tensorflow's official speech_commands. The speech recognition model is used to perform recognition training on 10 groups of speech files that have been classified by labels. The label of each group of speech files is a corresponding English word, and each group of speech files includes speech files in which the English word is read out by different speakers.

3. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 1, characterized in that: In step B, the individuals and numbers in the main population and the two auxiliary populations are the same, and the number of individuals in a single population is set to 20-40.

4. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 1, characterized in that: In the step B, when adding Gaussian noise, each bit element of the original audio sample in the form of a binary string is traversed, and the currently traversed element is converted from 1 to 0 or from 0 to 1 with a set conversion probability.

5. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 1, characterized in that: The step F comprises the following specific steps: F1: Using the elitist method, the first P individuals of the sorted contemporary individuals in the main population and the two auxiliary populations are selected as elite individuals with a set retention probability, and the remaining NP individuals are selected as non-elite individuals; F2: For the main population and two auxiliary populations composed of contemporary individuals, the average crossover method is used to perform crossover operations, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the crossover operation; then enter step F3; F3: Perform mutation operation on the main population and two auxiliary populations that have completed the crossover operation, wherein the elite individuals in the main population and the two auxiliary populations do not participate in the mutation operation; then enter step F4; F4: Using the method in step D, first calculate the fitness score of each offspring individual in the main population and the two auxiliary populations that have completed the genetic operation; then sort all offspring individuals in the main population and the two auxiliary populations that have completed the previous round of genetic operations in order of fitness scores from low to high; then, according to the set optimal individual selection threshold, the first L individuals in each auxiliary population are sorted as the optimal individuals, and the 2L optimal individuals are added to the main population that has completed the genetic operation; then, in order of fitness scores from low to high, all offspring individuals in the main population that has been added with the 2L optimal individuals are sorted, and the 2L individuals with the highest fitness scores that are ranked last in the main population are eliminated; finally, the main population with the 2L offspring individuals eliminated is used as the main population that has completed the genetic operation and individual update; then return to D.

6. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 1, characterized in that: In the step F, the average crossover method is used to perform the crossover operation; when performing the genetic operation, a three-population parallel genetic algorithm is used; a main population and two auxiliary populations are independent of each other, different populations perform different genetic operations, and no information exchange is performed between the three populations; for the main population and the two auxiliary populations, the genetic operation is performed after setting different genetic operation parameters.

7. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 6, characterized in that: In the step F, the first auxiliary population is set to have a small mutation probability and a large crossover probability, that is, the mutation probability of the first auxiliary population is the minimum mutation probability among the three populations, and the crossover probability of the first auxiliary population is the maximum crossover probability among the three populations; the second auxiliary population is set to have a small crossover probability and a large mutation probability, that is, the crossover probability of the second auxiliary population is the minimum crossover probability among the three populations, and the mutation probability of the second auxiliary population is the maximum mutation probability among the three populations.

8. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 5, characterized in that: In the step F1, when performing a crossover operation on the population, two contemporary individuals A and B are randomly selected from the non-elite individuals in the population as parent individuals, and then each element in the contemporary individuals A and B is traversed, and the elements at the corresponding positions of the contemporary individuals A and B are cross-exchanged according to the set population crossover probability threshold to complete the crossover operation of the population and obtain the contemporary non-elite individuals generated after the population crossover.

9. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 5, characterized in that: In step F3, when performing mutation operation on the population, for each contemporary non-elite individual generated after the crossover of the population, each element of the contemporary non-elite individual is traversed, and through the set population mutation threshold, some elements in the contemporary non-elite individual are mutated to complete the population mutation operation, and obtain the population that has completed the genetic operation, and the offspring individuals corresponding to each population; the offspring individuals in the main population that has completed the genetic operation and the two auxiliary populations are composed of elite individuals that have not undergone crossover mutation operations and non-elite individuals that have undergone crossover mutation operations.

10. The method for generating adversarial audio samples based on three-population parallel genetic algorithm according to claim 7, characterized in that: Main population crossover probability threshold M C The range is 40%≤M C ≤60%; First auxiliary population crossover probability threshold A 1C The range is 60%<A 1C ≤90%; Second auxiliary population crossover probability threshold A 2C The range is 10% ≤ A 2C <40%; Among them, M C M is the abbreviation of main, A 1C and A 2C The A in the figure is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population, and the c in the subscripts are the abbreviations of crossover; Main population mutation probability threshold M m The range is 0.0004≤M m ≤0.0006; first auxiliary population mutation probability threshold A 1m The range is 0.0001≤A 1m <0.0004; The second auxiliary population mutation probability threshold A 2m The range is 0.0006<A 2m ≤0.0009; where M m M is the abbreviation of main, A 1m and A 2m The A in the expression is the abbreviation of auxiliary, the 1 and 2 in the subscripts represent the first auxiliary population and the second auxiliary population respectively, and the m in the subscripts are the abbreviations of mutation.