Shilling Attacks And Detection In Social Recommender Systems

With the rapid development of electronic commerce and the rise of online social networks, social recommender systems using social relationships as an additional input have become a new research direction. Social recommender systems can make use of the similarity reflecting by user relations, so they play an important role in solving the problem of cold start and improving the accuracy of traditional recommender systems. However, open nature of social recommender systems makes them easy to be influenced by spam information(spam ratings and relationships) injected by spammers. This problem is denoted by “shilling attacks” and it seriously affects the fairness and authenticity of recommendation results, thus, reduces the trust of users to systems.Social recommender systems are seen as the combination of traditional recommender systems and online social networks. Most of the existing studies focus on shilling attack detection in rating driven recommender systems or relationship driven social network, but less attention has been paid to shilling attacks in social recommender systems that are driven by ratings and relationships both. To solve the deficiency of existing research, this paper first models shilling attacks in social recommender systems, then analyzes the feature extraction methods in detecting spam information in recommender systems and online social networks respectively, finally the shilling attack detection technology in social recommender systems is studied. This paper conducts the following aspects of research:(1) To study shilling attack models in social recommender systems, and these models are analyzed from the view of the attack cost and effect. By analyzing the principle of existing social recommendation algorithms, the possible attack behavior of spammers can be summarized, thus, the shilling models can be proposed. Then, through analyzing the impact of the attack models to recommendation results, the attack effect of the proposed shilling attack models can be obtained.(2) To study shilling attacks in recommendation systems, and put forward a novel detection method based on popularity features. Traditional methods detect shilling attackers from the difference of rating patterns between normal users and attackers, but these methods are hard to detect new forms of shilling attacks. To solve this problem, choice behavior of attackers and normal users is analysis, in order to obtain the popularity distribution difference of user profiles between normal users and attackers. Therefore, popularity distributions of user profiles are applied to derive features for classifiers to detect shilling attacks in recommender systems.(3) To study shilling attacks in online social networks, and put forward a novel shilling attack detection method based on Laplace score. Shilling attackers in online social networks promote their popularity through inserting bogus relationships to deliver social spams in networks. Traditional methods train models using high dimension features, resulting in the low detection accuracy. To solve this problem, a unsupervised learning feature selection method is used. More specifically, Laplace score, which represents the ability to maintain local information of the original space, is used to select features. On this basis, a semi-supervised method is combined to detect shilling attacks in social networks.(4) To study shilling attacks in social recommender systems, and put forward a novel shilling attack method based on semi-supervised co-training. Users in social recommender systems include rating features and relationship features, which can be used to train classifiers on the two kinds of feature subsets synchronously. At the same time, considering the problem of insufficient labels in practice, semi-supervised co-training algorithm is used to construct the model to detect shilling attacks in social recommender systems.