Text backdoor attack method and device

Abstract

The invention provides a text backdoor attack method and device, and the method comprises the steps: obtaining a test sample; and inputting the test sample into a damage model of an injection back door to obtain a judgment result of the test sample, wherein the victim model of the inserted backdoor is obtained by the following method: determining a specific text style as a triggering feature of the backdoor attack, performing text style migration on a to-be-poisoned sample to obtain a training sample with the trigger feature, and training a damage model according to the training sample to obtain the damage model of the injection back door. Through the steps, the text backdoor attack technology can be researched, the safety and robustness of the natural language processing model can be detected, and the risk that the natural language processing model is put into practical application is managed and controlled.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a text backdoor attack method and device. Background technique [0002] Backdoor attacks are an emerging security threat to machine learning, especially deep learning models. Backdoor attacks usually inject the backdoor into the victim model during the training process, so that the victim model works normally when faced with normal input in the test phase, which is no different from the normal model without the backdoor, but when faced with the input containing pre-designed trigger features , which can be output according to the pre-design. For example, a face recognition system that has been attacked by a backdoor can correctly recognize general face images after deployment, but when it encounters a face wearing glasses of a preset specific color, it will be recognized as a specific face. person, regardless of which person the face with glasses corresponds to. [0003] Sinc...

AI Technical Summary

Problems solved by technology

[0004] In the prior art, there is a lack of research on the security and robustness of natural language processing models by backdoor attacks, resulting in the inability to control the risks of putting natural language processing models into practical applications in advance

Images