Research On Multi-Granularity Textual Backdoor Attack Method

The backdoor attack of the deep neural network is a malicious attack that adds triggers to the training data during model training to embed a backdoor in the model.Usually,after the trigger is poisoned on the original sample,the backdoor model will predict wrong malicious results.In addition,due to the complexity of text data,there are still certain defects in the current text backdoor attack,which leads to the problem of incoherent sentences when modifying the text,and because the text modification features are too simple and the trigger features are few,it will lead to poisoning Samples are easily inspected.Aiming at the above problems of text backdoor attack,this paper conducts research on how to modify the text to increase the complexity of the trigger.The main work is as follows.Analyze the attack forms and problems of text backdoor attacks,and propose an overall scheme of multi-granularity textual backdoor attacks on the existing problems at the present stage,and propose a multi-granularity modification generation algorithm for text,using two fine-grained upper the modified model rewrites the original sentence,so that the modified text can get the optimal candidate multi-granularity modified text.In order to make the text backdoor attack successful,it is proposed to use the samples with multi-granularity modification characteristics obtained by the multi-granularity text modification generation algorithm to establish the backdoor model,and add the concept of opposite samples to the model training to fine-tune the backdoor model to ensure the backdoor model While being able to be stably activated by samples containing multi-granularity triggers,it can also maintain a high accuracy rate of clean samples.In this paper,on two data sets of classification tasks widely used in the real world,three victim models are used to carry out multi-granularity textual backdoor attack experiments under undefended and defensive strategies respectively,and a detailed comparison with various baseline method models For comparison and evaluation,the experiment proves that the multi-granularity text backdoor attack can achieve excellent performance.