一种探测推荐系统托攻击的无监督算法

An Unsupervised Algorithm for Detecting Shilling Attacks on Recommender Systems

托攻击是当前推荐系统面临的重大安全性问题之一.开发托攻击探测算法已成为保障推荐系统准确性与鲁棒性的关键.针对现有托攻击探测算法无监督程度较低的局限,在引入攻击概貌群体效应的定量度量及基于此的遗传优化目标函数的基础上,将自适应参数的后验推断与攻击探测过程相融合,提出了迭代贝叶斯推断遗传探测算法,降低了算法探测性能对系统相关先验知识的依赖.实验结果显示这种算法能够有效探测各种常见攻击.

Shilling attack is one of the significant security problems involved in recommender systems. Developing detection algorithms against shilling attacks has become the key to guaranteeing both the preciseness and robustness of recommender systems. Considering the low degree of unsupervised features the existing algorithms suffer from, this paper proposes an iterative Bayesian inference genetic detection algorithm （IBIGDA） through the introduction of the quantitative metric for the group effect of attack profiles and the corresponding object function for genetic optimization. This algorithm combines the posterior inference for the adaptive parameters with the process of attack detection, thus relaxes the dependence of the detection performance on the relating prior knowledge of the systems. Experimental results show that this algorithm can effectively detect shilling attacks of typical types.