Adversarial sample generation method, system and device for outlier removal method

Abstract

The invention belongs to the field of image recognition, particularly relates to an adversarial sample generation method, system and device for an outlier removal method, and aims to solve the problems that an adversarial sample adopted by existing classification model training based on deep learning cannot make an image classification error under an outlier removal method; and therefore, the trained classification model is poor in robustness and low in accuracy. The method comprises the following steps: acquiring a training data set with category labels, inputting three-dimensional point cloud data into a classification model, calculating classification loss, respectively calculating the gradient of the classification loss relative to the three-dimensional point cloud data and the gradient of the classification loss relative to outlier-removed three-dimensional point cloud data, and fusing the two gradients by multiplying a scaling factor to generate fusion disturbance, and applying the fusion disturbance to the three-dimensional point cloud data for repeated iteration to generate an adversarial sample. The generated adversarial samples can still cause image classification errorsunder the condition that outliers are removed, and the robustness and classification accuracy of the trained model are improved.

Description

technical field [0001] The invention belongs to the field of image recognition, and in particular relates to an adversarial sample generation method, system and device for outlier removal methods. Background technique [0002] The 3D point cloud classification model based on deep learning has been greatly developed in recent years, and can achieve a classification accuracy of about 90% on the ModelNet40 dataset. But at the same time, in the field of image classification, the classification model based on deep learning is vulnerable to the attack of adversarial samples. From the appearance, the adversarial samples are almost the same as the normal samples, but the adversarial samples will make the classification model make mistakes. However, there are few studies on adversarial examples in the field of 3D point cloud, and the field of 3D point cloud classification is related to engineering projects with high security requirements such as automatic driving and machine grasping...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the above-mentioned problems in the prior art, the existing image classification model based on deep learning adopts conventional adversarial examples for training, which has poor robustness and cannot me

Images