基于特征子集的推荐系统托攻击无监督检测

摘要

To solve the problem that existing recommender systems based on collaborative filtering are vulnerable to the shilling attack, this paper proposes an Unsupervised Detection Algorithm of Shilling Attack Based on Feature Subset(UnDSA-FS). A feature named Kurtosis Coefficient of Interest(KCI) is proposed to describe the intensity degree of user’s interest. Taking the KCI and other existed features as candidate feature set, this algorithm uses unsupervised feature selection method to choose proper feature subset for different attack strategies. It computes the distance sum of each user, sorts the users by the distance sum and identifies the attack target. It sets a sliding window on the sorted user sequence, and filters the attack users by calculating the mean rating deviation of attack target. Experimental result verifies that the information gain of KCI is higher than existing features’, and the proposed UnDSA-FS has a better performance in stability and precision compared with existing unsupervised detection methods.%针对现有基于协同过滤的推荐系统易受托攻击影响的问题，提出一种基于特征子集的推荐系统托攻击无监督检测算法。利用现有攻击模型在项目选择上的随机性，给出一种描述用户兴趣集中程度的特征属性：兴趣峰度系数。将该系数与已有的推荐系统用户特征属性结合作为备选特征集，采用无监督特征选择方法为不同类型托攻击选取相应的检测特征子集。根据选择出的特征子集计算每个用户的离群度，以此进行排序并确定攻击目标，在已排序的用户序列上设置滑动窗口，通过计算窗口内攻击目标的平均评分偏移值对攻击用户进行过滤。实验结果证明，兴趣峰度系数的信息增益高于已有的特征属性，基于特征子集的无监督检测算法相比于现有的无监督检测方法具有更高的稳定性和精准度。