基于非对称半监督集成SVM的托攻击检测方法

Shilling Attack Detection Approach Based on Asymmetric Semi-supervised Ensemble SVM

受推荐系统在电子商务领域重大经济利益的驱动,恶意用户以非法牟利为目的实施托攻击,操纵改变推荐结果,使推荐系统面临严峻的信息安全威胁,如何识别和检测托攻击成为保障推荐系统信息安全的关键。传统支持向量机(SVM)方法同时受到小样本和数据不均衡两个问题的制约。为此,提出一种半监督SVM和非对称集成策略相结合的托攻击检测方法。首先训练初始SVM,然后引入K最近邻法优化分类面附近样本的标记质量,利用标记数据和未标记数据的混合样本集减少对标记数据的需求。最后,设计一种非对称加权集成策略,重点关注攻击样本的分类准确率,降低集成分类器对数据不均衡的敏感性。实验结果表明,本文方法有效地解决了小样本问题和数据不均衡分布问题,获得了较好的检测效果。

Driven by the significant economic benefits of the recommendation system in the field of electronic commerce, malicious users implement shilling attacks for the purpose of illegal profit making, manipulate and change the recommendation results. Recommendation system faces major security threats, so how to detect shilling attacks becomes crucial to guarantee the security of recommender system. Shilling attack detection method based on traditional support vector machine is challenged by the small sample problem and the asymmetric distribution problems. This paper proposes an asymmetric ensemble method with semi-supervised SVM for detecting shilling attacks to address these problems. In the first stage, I use support vector machine （SVM） as the basic classifier, and in the second stage us KNN classifier to improve the quality of the new training examples which are transformed from the boundary vectors, and unlabeled data are presented to produce a number of semi-supervised classifiers. Then by the asymmetric semi-supervised ensemble strategies, these classifiers are combined . The ensemble classification model pays more attention to the positive samples than the negative ones. The proposed approaeh is applied to benchmark problems, and the simulation results show its validity.