Profile Injection Attack Detection in Recommender System

Abstract

Recommender systems are backbone for ecommerce website today. It is not restricted to ecommerce websites; social networking websites are also highly dependent on recommender systems. These are using recommender systems for providing personalized services to their users or customers. Recommender systems are desired to attain a high level of accuracy while making the predictions which are relevant to the customer, as it becomes a very tedious task to explore the thousands of relevant items from the huge database. Different ecommerce companies use different types of recommender systems based on their requirement. But a threat comes in frame in last one decade i.e. profile injection in the system to affect the accuracy of the collaborative recommender system. These attacks degrade the accuracy of recommendations. In this thesis, we focused on the behavior of attacks in the system. When attacks are inserted some change comes in the recommendations, we measure these changes in the form of prediction shift. This prediction shift may be positive or negative. In the next phase we focused on the detection of these attacks so that these attacks no longer affect the accuracy of the system. For this purpose, we use machine learning models and detection attributes. In the supervised profile classification approach, we compare six models and compare their accuracy. Although, none of the model gives 100% accuracy but some models gives good accuracy. We pick best performer models and proposed our ensemble model by using voting technique of ensembling. Although our proposed model doesn’t give best accuracy but it will never give worst accuracy in any case. Next we give a comparative study of unsupervised models.