针对商品推荐系统的混淆托攻击半监督检测研究

利用标记用户，提出一种半监督检测算法，对推荐系统混淆托攻击进行有效检测。为降低评分偏移策略对检测的影响，改进K均值聚类算法（K-Means Clustering Algorithm, K-Means）算法，提出共同关注项；为凸显噪音注入策略对检测的影响，过滤共同关注项，计算标记用户特征指标，利用二次特征提取的方法，给出区分用户概貌的先验知识。在此基础上压缩特征指标向量，使攻击概貌和普通概貌富集到空间两端，最终识别攻击用户。比较半监督检测、主成分分析和贝叶斯检测算法，结果表明，该检测算法对混淆托攻击在不同填充率下有较高的检测准确率。

Semi-Supervised Detection Method of Confused Shilling Attack for Commodity Recommendation System

Tag user is used, a semi-supervised detection algorithm is proposed to effectively detect the recommended system confusion shilling attack. In order to reduce the impact of the scoring offset strategy on detection, the K-Means algorithm is improved to propose common concerns; In order to highlight the impact of noise injection strategies on detection, the common interest items are filtered, the tag user feature indicators are calculated, and then the secondary feature extraction method is used to provide prior knowledge for distinguishing user profiles. On this basis, the feature index vector is compressed to enrich the attack profile and common profile to both ends of the space. Finally, the attack user is identified. Semi-supervised Detection, Principal Component Analysis, Bayesian prediction are compared. Experiments show that the detection algorithm has a high detection accuracy for different confused attacks at different filling rates.