Multi-priori-based black box adversarial test sample generation method and device

Abstract

The invention discloses a multi-prior-based black box adversarial test sample generation method and device. The method comprises the steps of setting a plurality of hyper-parameters required for adversarial sample generation; initializing the adversarial sample and starting iteration; obtaining a plurality of different priors to obtain a group of orthogonal bases; estimating the similarity between the real gradient and each orthogonal vector; optimizing the objective function, and minimizing the expected difference between the estimated gradient and the real gradient; sampling a plurality of random vectors; estimating gradients according to a stochastic gradient estimation method. Therefore, the attack success rate of the neural network can be improved, or under the condition of the same attack success rate, the sampling frequency for estimating the gradient is reduced, and the generation of adversarial samples is accelerated.

Description

technical field [0001] The invention relates to the field of machine learning, in particular to a method and device for generating black-box adversarial test samples based on multiple priors. Background technique [0002] Deep learning has achieved groundbreaking progress in many fields, but recent studies have shown that they are vulnerable to adversarial examples. The so-called adversarial example refers to adding some small perturbations invisible to the naked eye to the normal picture, so that the deep learning model that can work correctly can make prediction errors. The generation methods of adversarial examples include white-box methods and black-box methods. The white-box method needs to assume that we are fully aware of the information of the attacked deep learning model, so as to use the algorithm based on gradient ascent to generate adversarial samples. Black-box methods are mainly divided into score-based black-box methods, decision-based black-box methods, and...

AI Technical Summary

Problems solved by technology

[0002] Deep learning has achieved groundbreaking progress in many fields, but recent studies have shown that they are vulnerable to adversarial examples

Images