【作者】 刘子豪； 庄毅；

【Author】 Liu Zihao and Zhuang Yi (College of Information Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 210016)

【机构】 南京航空航天大学信息科学与技术学院；

【摘要】 针对目前电子邮件安全网关不能很好地支持敏感信息检测问题,深入研究了Winnow算法和Markov模型,在N-Gram语言模型的基础上,提出了一种邮件特征选择方法——Markov-Gram,该方法以句子为单位进行特征项的选取,不仅保留了更多的语义信息,而且可以有效地减少特征项的数目,解决"维度灾难"问题;提出一种Winnow算法训练过程中初始权重生成方法,该方法融入了电子邮件结构特点以及关键词的密级等相关信息,可以有效减少Winnow算法训练的轮数;基于以上两点提出了电子邮件敏感信息检测算法——EWMG;实验结果表明:提出的上述两种方法效果很好,采用EWMG算法对邮件敏感信息检测时,其正确率较Winnow算法有明显提高.更多还原

【Abstract】 Nowadays,the Email secure gateway can not provide good support to the function of sensitive information detection.After deeply researching into the Winnow algorithm and Markov model,the authors put forward an Email feature selection method,the Markov-Gram feature selection method.The method regards each sentence in an Email as a candidate feature instead of the words,by way of which it can preserve more semantic information and the dimensionality of feature denotations in an Email can be decreased.A vector initial method is proposed,which is integrated with the characteristic of Email and the secrete level of key words,used by the winnow algorithm training process to decrease the training times.Based on the two points above,a sensitive Email detection training method-EWMG,is proposed.Finally,the experiment proves that the methods proposed in this paper is better than the normal Winnow detect algorithm.更多还原