Abstract: Because of the open nature of collaborative recommender systems and their reliance on user-specified judgments for building profiles, an attacker could affect the prediction by injecting a lot of biased data. In order to keep the authenticity of recommendations, the attack detection method based on Statistical Process Control (SPC) was proposed. The method constructed the Shewhart control chart by using the users' deviation from the average of rating numbers and detected attackers according to the warning rules of the chart, thus improving the robustness of collaborative recommender systems. The experiments demonstrate that the method is effective with high precision and high recall against a variety of attack models.

Key words: collaborative recommender system, Statistical Process Control (SPC), deviation from average of rating number, shilling attack, attack detection

摘要： 针对恶意攻击者利用协同推荐系统用户偏好敏感的缺陷向系统中注入虚假数据破坏推荐结果真实性的问题,提出基于统计过程控制(SPC)的协同推荐攻击检测方法。该方法将用户概貌项目评价数偏离度作为服务质量控制属性构建休哈特控制图,利用判异规则检测攻击用户,从而完善协同推荐系统模型。实验证明这种检测方法对各种不同的攻击模型都有较高的检测准确率和查全率。

关键词: 协同推荐系统, 统计过程控制, 用户概貌项目评价数偏离度, 托攻击, 攻击检测