A Black-box Adversarial Sample Generation Algorithm Based on Hidden Space Clustering

Abstract

The invention discloses a black-box adversarial sample generation algorithm based on latent space clustering, comprising the following steps: step 1, using a convolutional self-encoder to extract the feature representation of an image sample; step 2, when the attacked model needs to be misclassified When , use the feature representation of the image sample to perform the generation algorithm of the misclassified adversarial sample to obtain the adversarial sample; step 3, when it is necessary to carry out target confrontation against the attacked model, use the feature representation of the image sample to perform the target adversarial sample generation algorithm to obtain the adversarial sample . The present invention uses a convolutional self-encoder to extract feature representations of image samples, thereby realizing hidden space clustering, and realizing a black-box confrontation sample generation algorithm without knowing the structure of a target model.

Description

technical field [0001] The invention relates to the fields of machine learning and pattern recognition, in particular to a black-box adversarial sample generation algorithm based on latent space clustering. Background technique [0002] In recent years, Deep Neural Network (DNN) has brought revolutionary changes in the field of machine learning and pattern recognition. DNN has achieved a lot of results on many pattern recognition tasks (especially visual classification problems), and its excellent performance has also attracted extensive attention in many other fields. In the field of adversarial sample attacks, researchers have also proposed many DNN-based adversarial sample generation algorithms. However, a large number of experiments have shown that there are some loopholes in the DNN-based algorithm, and there are currently three main problems: [0003] 1. Most algorithms need to obtain the classification boundaries in the feature space through training samples, which ...

AI Technical Summary

Problems solved by technology

[0005] 3. Most algorithms only consider the classification between different categories, and do not consider the multimodal distribution of data within the same category.

Images