基于协同过滤的电子商务推荐系统极易受到托攻击,托攻击者注入伪造的用户模型增加或减少目标对象的推荐频率,如何检测托攻击是目前推荐系统领域的热点研究课题.分析五种类型托攻击对不同协同过滤算法产生的危害性,提出一种特征选择算法,为不同类型托攻击选取有效的检测指标.基于选择出的指标,提出两种基于监督学习的托攻击检测算法,第一种算法基于朴素贝叶斯分类；第二种算法基于k近邻分类.最后,通过实验验证了特征选择算法的有效性,及两种算法的灵敏性和特效性.%Most of the e-business recommender systems are based upon collaborative filtering (CF) algorithms. Since such systems have been shown to be vulnerable to shilling attacks in which malicious user profiles ate inserted into the system in order to push or nuke the predictions of some targeted items, shilling attack detection has recently become a hot research topic in recommender systems. Firstly, the effectiveness of five types of attacks against different CF algorithms is analyzed. Secondly, a feature selection algorithm is presented. Two kinds of shilling attack detection algorithms based on supervised learning are then proposed: the first one is based on naive Bayesian classifier,and the second one is based on k nearest neighbor (κNN) classifier. At last,experimental results show the effectiveness of the feature selection algorithm and the sensitivity and specificity of these two kinds of detection algorithms.
展开▼
