单分类人机检测方法

Human-Bot Detection with One-Class Classifiers

针对人机检测领域中脚本机器人种类多样的问题,提出了一种利用单分类方法基于鼠标行为的人机检测技术,实现对带有行为模拟功能的高级脚本机器人的检测。利用JavaScript获取网页页面正常用户的鼠标行为数据,经过数据预处理、特征提取过程获得描述用户行为的特征向量并送入单分类器中,建立人机检测模型。为验证单分类方法的有效性,通过统计分析网页页面正常用户的鼠标行为数据,设计3种具有模拟人类功能的脚本机器人,模拟正常用户完成网页页面登录场景任务。使用直线类型的脚本机器人样本参与特征选择,之后提取特征子集,完成单分类器在人机检测领域的性能评价。实验结果表明:单分类器对直线、规则曲线、不规则曲线3种脚本机器人检测效果最好的等错误率分别为7.25%、9.45%和12.7%;较二分类方法对未知类型脚本机器人检测效果最好的等错误率为20.56%;单分类方法具有良好的泛化性。

A behavior-based human-bot detection technology is proposed to address the problem of a variety of scripted bots in the field of human-bot detection,where one-class classifier is considered to establish a human mouse behavior model to detect sophisticated scripted bots with behavioral simulation functions.JavaScript is installed on the web page to collect the normal users’mouse behaviors,the feature vectors describing the users’behaviors are obtained and sent to the one-class classifiers to construct the human-bot detection model after the data preprocessing and feature extraction process.Statistically analyzing the mouse behaviors of normal users on the web,three types of scripted bots with behavior simulation function are designed to simulate the normal user to complete the task of logging into the web.According to the feature subset selected by the linear scripted bots participated in the feature selection process,the performances of one-class classifiers in human-bot detection are evaluated.Experimental results show that the best equal error rates of one-class classifiers for linear,regular,and irregular curves scripted bots reach 7.25%,9.45%,and 12.7%,respectively.Compared with two-class method whose best equal error rate is 20.56%,the one-class method is endowed with excellent generalization property for unknown scripted bots detection.