对抗环境下的线性分类器对抗性比较

摘要

Machine learning algorithms provide a well solution for many security applications. Machine learning algorithms themselves , however, face the thread of adversary attack. In order to analyze the impact of adversary attacks on machine learning algorithms , the paper presents an adversary attack model in line with some actual situations and compares the antagonism of some linear classifier under this model. The performances of these adversarial classifiers are evaluated on a large public spam corpus. The experiment results show that SVM is more antagonism than other linear classifiers.%机器学习算法为很多安全应用提供了良好的解决方案,然而机器学习算法本身却面临被敌手攻击的威胁.为分析敌手攻击对机器学习算法造成的影响,本文提出符合某些特定场合的敌手攻击模型,并在该模型下比较几种线性分类器的对抗性.最后在垃圾邮件过滤公开数据库上进行测试,实验结果表明,支持向量分类器具有相对较好的对抗性.