Voiceprint recognition poisoning sample generation method based on migration training

Abstract

The invention discloses a voiceprint recognition poisoning sample generation method based on migration training. The voiceprint recognition poisoning sample generation method comprises the following steps: (1) preprocessing a voice data set; (2) building a voiceprint recognition model; (3) obtaining a feature representation space of the migration training task data set; (4) selecting a target sample and a base sample from the test set; (5) generating a poisoning sample by using an optimization algorithm; and (6) adding the poisoning sample into the original training set for migration training: under the condition of not changing the weight of the original model, only retraining the replaced softmax layer to be suitable for the current classification task, and adding one poisoning sample into the original training set by the training set. In a test phase, a target sample will be misclassified as a tag of a base sample. According to the invention, the attack success rate is greatly improved.

Description

technical field [0001] The invention relates to a method for generating poisoned samples for voiceprint recognition based on migration training, and the invention belongs to the field of deep learning security. Background technique [0002] Deep learning is one of the most common technologies of artificial intelligence. It has developed rapidly in recent years. It can handle many complex tasks, including image recognition, object detection, speech recognition, signal processing, etc. Voiceprint recognition is one of the most mature biometric authentication technologies at present. The voiceprint recognition technology based on deep learning has greatly improved its accuracy. But it turns out that the deep learning model is vulnerable to attacks. Attackers discover the weaknesses of the model and create adversarial samples that are different from the original samples, so that the trained model cannot run correctly. Therefore, the attack method for voiceprint recognition has a...

AI Technical Summary

Problems solved by technology

In previous work, poisoning attacks usually change part of the class labels of the training data to poison the model training, but this will not only greatly reduce the accuracy of the test set, but also classify the specified samples without distinction, making this attack method practical not tall

Images