攻击标签信息的堆栈式支持向量机  

Stacked Support Vector Machine Based on Attacks on Labels of Data Samples

作　　者：金耀[1] 徐丽亚[1] 吕慧琳 顾苏杭 JIN Yao;XU Li-ya;LV Hui-lin;GU Su-hang(School of Information Science and Engineering,Changzhou University,Changzhou,Jiangsu 213164,China;School of Digital Media,Jiangnan University,Wuxi,Jiangsu 214122,China;College of Information Engineering and Technology,Changzhou Vocational Institute of Light Industry,Changzhou,Jiangsu 213164,China)

机构地区：[1]常州大学信息科学与工程学院,江苏常州213164 [2]江南大学数字媒体学院,江苏无锡214122 [3]常州轻工职业技术学院信息工程学院,江苏常州213164

出　　处：《计算机科学》2020年第1期110-116,共7页Computer Science

基　　金：国家自然科学基金(81701793);常州市科技计划项目(CJ20190016)~~

摘　　要：真实数据集中存在的对抗样本易导致分类器取得较差的分类性能,但如果其能够被合理利用,分类器的泛化能力将得到显著提高。针对现有大部分分类器并没有涉及对抗样本信息的问题,提出一种攻击标签信息的堆栈式支持向量机。该方法从给定的初始数据集中选取一定比例的样本,并攻击所选取样本的标签,使之成为对抗样本,即将样本标签替换成其他不同类型的标签,利用支持向量机训练包含对抗样本的数据集,从而生成对抗支持向量机。计算对抗支持向量机的输出误差相对于输入样本的一阶梯度信息,并将其嵌入到输入样本特征中以更新输入样本。将更新后的样本输入到下一个对抗支持向量机中,并重新训练。以堆栈方式级联一定数目的对抗支持向量机,直至取得最好的分类性能。原理分析与实验结果表明,基于对抗样本的一阶梯度信息不仅提供了分类器输出与输入之间的一种正相关关系,而且为堆栈式支持向量机中的子分类器提供了一种新的堆栈方式,并提高了分类器的整体性能。As for the adversarial data samples which indeed exist in real-world datasets,they can mislead data classifiers into correct predictions which results in poor classification.However,reasonable utilization of the adversarial data samples can distinctly improve the generalization of data classifiers.Since most of existing classifiers do not take the information about adversarial data samples into account to build corresponding classification models,a stacked support vector machine called S-SVM based on attacks on the labels of data samples which aims to obtain outperformed classification performance by learning the adversarial data samples was proposed.In a given dataset,a certain percentage of data samples are randomly chosen as adversarial data samples,in other words,the labels of these chosen data samples are substituted by the other labels included in the given dataset which are different from the original labels of the chosen data samples.Adversarial support vector machine(A-SVM)can be consequently generated by using the support vector machine(SVM)to train the given dataset which contains the adversarial data samples.The first-order gradient information on the output error of the generated A-SVM with respect to the input samples can be then computed,and the input samples will be updated by embedding the first-order gradient information into the original feature space of the input samples.Consequently,the updated data samples can be input into next A-SVM to be trained again to gradually improve the classification performance of the current A-SVM.As a result,S-SVM is formulated by stacking some A-SVMs layer by layer,the best classification results can also be obtained by the corresponding S-SVM.In terms of theoretical analysis and experimental results on UCI and KEEL real-world datasets,the mathematically computed first-order gradient information based on learning the adversarial data samples not only provide a positive relation between the outputs and the inputs of a classifier,but also indeed provide a novel w

关 键 词：堆栈结构 对抗样本 标签攻击 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]