Attack Detection in Recommender Systems using Clustering Techniques

by Rathee, Nupur

Abstract (Summary)
Expansion of internet has provided people with a plethora of information and choices. It has also brought along with it a problem to sort the abundant available information and to dig out pieces that are useful to us. For these reasons, people have always relied on recommendations by friends and family because they trust them. Recommender systems try to imitate this process of seeking recommendations from people. Collaborative recommender systems are social systems. They provide their users with predictions and recommendations, which are based upon their past opinions and account details. Today, they have become an important tool in e-commerce. They are heavily utilized on the web to filter information and generate recommendations for the people who use commercial service providers like Amazon, eBay, Yahoo! and many others. A user-based collaborative recommender system works by collecting profiles of users with similar tastes i.e. which have a similar rating pattern and then uses these ratings to generate recommendation for such people who have similar opinions and preferences.   Recommender systems are entirely based on the input provided by the users or customers. Due to this reason, they tend to become highly vulnerable to outside attacks. The product sellers who are interested in popularizing their own product to generate more revenue or else interested in making the product of their opponent less popular in the customer space, might be interested in biasing these recommender system, which have an influence on the customers. Such attackers can use automated tools to create and throw fake profiles in the recommender database, which either may rate their item high or may rate their opponent's item low. These collaborative recommender systems must always be open to users, in order to get their opinions. This is the reason why designing an attack-proof system is a complicated task. We aim to detect such attack profiles by looking at the rating pattern in the database, clustering the users to group together similar attack profiles and then identify and eliminate the attack groups. In this thesis, we analyze the performance of various clustering techniques to detect such attacks.
Bibliographical Information:


School:University of Cincinnati

School Location:USA - Ohio

Source Type:Master's Thesis



Date of Publication:01/01/2008

© 2009 All Rights Reserved.