A speech recognition model training method against adversarial sample attack

Abstract

The invention discloses a speech recognition model training method against adversarial sample attack, comprising the following steps: A: selecting a data set composed of speech files, B: sampling the speech files to obtain the collection tensor; C: collecting the collection tensor Extract the feature tensor; D: Input the feature tensor, the Chinese characters corresponding to the voice file, and the training parameters into the convolutional neural network model, and use the CTC function to map the output tensor to a different output at each time step in combination with the acoustic model and the language model. The probability of Chinese characters, and then select the maximum value of the probability to transcribe the corresponding Chinese characters as the actual output value; finally calculate the deviation value between the target value and the actual output value and record it as loss; E: Use the backpropagation method to update the convolutional neural network model The weight parameters of , update the training parameters of the input convolutional neural network model while updating the weight parameters. The invention can reduce the amount of training calculation and training time, and improve the ability of resisting malicious attacks.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to a speech recognition model training method against adversarial sample attacks. Background technique [0002] With the development of deep learning and neural network research, the application scenarios of automatic identification are becoming more and more extensive. In the field of speech, deep learning is reshaping the way we interact with computers, such as the personal assistants widely used in smartphones (Apple's Siri, Google's Assistant). These systems run speech recognition models to recognize and execute user commands. In fact, the research on automatic speech recognition is earlier than the appearance of computer, and the technology of speech synthesis and recognition can be traced back to the original vocoder. It was not until the new century that the research on artificial intelligence networks emerged in the field of speech recognition. Most artificial intelligenc...

AI Technical Summary

Problems solved by technology

[0003] While recent studies have shown that neural networks are vulnerable to adversarial attacks

This problem also exists in the field of speech recognition. The attacker adds a slight disturbance to the audio, which will cause the neural network to input completely different values, but the human ear cannot recognize the slight disturbance.

As the voice confrontation attacks that have appeared in recent years have become more and more aggressive and more and more types, the security issues in the field of voice recognition have become prominent, and it has also laid certain hidden dangers for the large-scale commercialization of voice recognition technology.

Images