Security enhanced sentence similarity computing model based on convolutional neural network

Abstract

Deep learning model shows great advantages in various fields. However, researchers pay attention to how to improve the accuracy of the model, while ignoring the security considerations. The problem of controlling the judgment result of deep learning model by attack examples and then affecting the system decision-making is gradually exposed. In order to improve the security of sentence similarity analysis model, we propose a convolution neural network model based on attention mechanism. First of all, the mutual information between sentences is correlated by attention weighting. Then, it is input into improved convolutional neural network. In addition, we add attack examples to the input, which is generated by the firefly algorithm. In the attack example, we replace the words in the sentence to some extent, which results in the adversarial data with great semantic change but slight sentence structure change. To a certain extent, the addition of attack example increases the ability of model to identify adversarial data and improves the robustness of the model. Experimental results show that the accuracy, recall rate and F1 value of the model are due to other baseline models.

This work was supported in part by the Major Scientific and Technological Projects of China National Petroleum Corporation (CNPC) under Grant ZD2019-183-006, in part by the Shandong Provincial Natural Science Foundation, China, under Grant ZR2020MF006, in part by the Fundamental Research Funds for the Central Universities of China University of Petroleum (East China) under Grant 20CX05017A, and in part by the Open Foundation of State Key Laboratory of Networking and Switching Technology (Beijing University of Posts and Telecommunications) under Grant SKLNST-2021-1-17.

Postprint (author's final draft)