Adversarial text sample generation method and device and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and provides an adversarial text sample generation method and device and a storage medium, the adversarial text sample generation method comprises the following steps: obtaining a first model obtained by training a plurality of first text samples; obtaining a plurality of second text samples which are similar to the first text samples in semantics; obtaining at least one to-be-trained text sample from the plurality of second text samples, and inputting the at least one to-be-trained text sample into the first model to obtain a prediction recognition result; determining at least one alternative text sample; performing parameter adjustment on the first model according to the at least one alternative text sample; and obtaining to-be-trained text samples from the remaining text samples of the plurality of second text samples until a prediction recognition result corresponding to each obtained to-be-trained text sample is consistent with an actual recognition result, and determining all alternative text samples as optimal confrontation text samples. By implementing the method, the optimal confrontation text sample with smooth semantics can be obtained, and the robustness of the model is improved.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence, and in particular to a method, device and storage medium for generating an adversarial text sample. Background technique [0002] Adversarial attacks are an emerging intersection of machine learning and computer security. The technology of adversarial attacks mainly includes two aspects: how to find the most effective and representative best adversarial samples, and the other direction is how to effectively defend against adversarial samples. There have been a lot of exploration and application of adversarial attacks in the field of machine vision. In the field of machine vision, the way to generate the best adversarial samples is often to add small perturbations to the original samples. After adding the perturbed samples, the human eye cannot see the difference. , but due to the addition of small perturbations, it often makes artificial intelligence (AI) model recognit...

AI Technical Summary

Problems solved by technology

However, different from the field of machine vision, the text samples in the field of natural language processing are usually discrete. If the best adversarial samples are generated by

Images