攻击标签信息的对抗分类算法
    点此下载全文
引用本文:陆兵1,顾苏杭1,2.攻击标签信息的对抗分类算法[J].计算技术与自动化,2019,(3):151-156
摘要点击次数: 59
全文下载次数: 0
作者单位
陆兵1,顾苏杭1,2 (1. 常州轻工职业技术学院 信息工程与技术学院江苏 常州 2131642. 江南大学 数字媒体学院江苏 无锡 214122) 
中文摘要:真实数据集中存在的对抗样本一方面易导致分类器取得较差分类结果,另一方面如果能够被合理利用,分类器的泛化能力将得到显著提高。针对现有大部分分类算法并没有利用对抗样本训练分类模型,提出一种攻击标签信息的对抗分类算法(ACA)。该方法从给定数据集中选取一定比例样本并攻击所选取的样本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(support vector machine,SVM)训练包含对抗样本的数据集,计算生成的SVM输出误差对于输入样本的一阶梯度信息并嵌入到输入样本特征中以更新输入样本。再次利用SVM训练更新后的样本以生成对抗的SVM(A-SVM)。原理分析与实验结果表明,一阶梯度信息不仅提供了一种分类器输出与输入之间的正相关关系,而且可提高A-SVM的实际分类性能
中文关键词:分类器  对抗样本  标签攻击  支持向量机
 
An Adversarial Classification Algorithm Based on Attacks on the Labels of Data Samples
Abstract:As for the adversarial data samples which indeed exist in real-world datasets,on the one hand,they can mislead data classifiers into correct predictions which results in poor classification. On the other hand,appropriate applications of the adversarial data samples can distinctly improve the generalization of data classifiers. However,most of existing classification methods do not take the adversarial data samples into account to build corresponding classification models. An adversarial classification algorithm (ACA) based on attacks on the labels of data samples which aims to obtain outperformed classification performance by learning the adversarial data samples is proposed. In a given dataset,a certain percentage of data samples are chosen as adversarial data samples,namely the labels of these chosen data samples are substituted by the other labels which are different from the original labels of the chosen data samples. A SVM model can be generated by using the support vector machine(SVM) algorithm to training the given dataset which contains the adversarial data samples. And the first-order gradient information on the output error of the generated SVM with respect to the input samples can be computed. The input samples can be updated by embedding the first-order gradient information into the original input samples. Consequently,adversarial SVM (A-SVM) can be generated by using the SVM alg-orithm again to train the updated input samples. In terms of theoretical analysis and experimental results on UCI real-world datasets,the mathematically computed first-order gradient information not only provided a positive relation between the outputs and the inputs of a classifier,but also indeed can improve the actual classification performance of A-SVM.
keywords:classifiers  adversarial data samples  attacks on labels  support vector machine(SVM)
查看全文   查看/发表评论   下载pdf阅读器