Unsupervised detection of obfuscated diverse attacks in recommender systems

Abstract

Unsupervised detection of obfuscated diverse attacks in recommender systems

Saad Sajid Hashmi Directed by Professor Sang-Wook Kim Department of Computer and Software Graduate School of Hanyang University

Recommender systems estimate ratings for the items that have not been seen by the user yet. The output of recommender systems depends on the ratings given in user profiles. Biased ratings of attack profiles have a significant impact on the effectiveness of collaborative recommender systems. Previous work has shown standard memory-based recommendation algorithms, such as the k-nearest neighbor (kNN)-based algorithm, susceptible to the attacks compared with model-based collaborative filtering (CF) algorithms. An obfuscated diverse attack strategy made model-based algorithms vulnerable to attacks. Attack profiles generated with this strategy are also able to avoid principal component analysis (PCA)-based detection. This thesis proposes an algorithm to detect obfuscated diverse attack profiles. Profiles’ pairwise covariance with each other is used to separate attack profiles from genuine profiles. Through extensive experiments, we demonstrate that the attack strategy is effective in both small and large datasets, and our algorithm detects these attack profiles with high accuracy.