Defending recommender systems by influence analysis

Abstract

Collaborative filtering (CF) is a popular method for personalizing product recommendations for e-commerce applications. In order to recommend a product to a user and predict that user's preference, CF utilizes product evaluation ratings of like-minded users. The process of finding like-minded users forms a social network among all users and each link between two users represents an implicit connection between them. Users having more connections with others are the most influential users. Attacking recommender systems is a new issue for these systems. Here, an attacker tries to manipulate a recommender system in order to change the recommendation output according to her wish. If an attacker succeeds, her profile is used over and over again by the recommender system, making her an influential user. In this study, we applied the established attack detection methods to the influential users, instead of the whole user set, to improve their attack detection performance. Experiments were conducted using the same settings previously used to test the established methods. The results showed that the proposed influence-based method had better detection performance and improved the stability of a recommender system for most attack scenarios. It performed considerably better than established detection methods for attacks that inserted low numbers of attack profiles (20-25 %).

Privacy-Aware Detection of Shilling Profiles on Arbitrarily Distributed Recommender Systems

“…The crucial point of the algorithm is HHT-based feature extraction method. To improve the success of attack detection, Morid et al [40] applied a k-NN supervised classification method on influential users, instead of the whole user set. Zhang and Zhou [41] propose an ensemble detection model (EDM) by introducing backpropagation neural network and ensemble learning technique to detect profile injection attacks through selecting and integrating parts of the base classifiers using voting strategy.…”

“…During the experiments, filler size parameter, I F , is varied from 3 to 60%, whereas attack size parameter, the number of inserted fake profiles, is kept constant at 1% since it is unreasonable to insert more in real-world cases [40]. Also, the attacked movies, i t , in the training set is chosen at random among the ones which have 80 to 100 ratings.…”

Multiview Ensemble Method for Detecting Shilling Attacks in Collaborative Recommender Systems

“…However, in the test sets, since the filler sizes of more than 95% genuine profiles are below 10%, the filler sizes are set to {2%, 4%, 6%, 8%, 10%}. In fact, if the filler size of shilling profiles is more than that of genuine profiles, the attacks can be easily detected by the features based on filler size, such as length variance (LengthVar) [5,6] and filler mean variance (FMV) [46]. Therefore, we set the maximum filler size in the test sets to 10%.…”

Defending Suspected Ratings in Collaborative Filtering Recommender Systems: A Fast Detection Method

“…Su et al [9] presented a spreading similarity algorithm to detect groups of similar attackers. Then, developed different features extracted from user profiles to capture attackers by exploiting classification-based methods [2,4,12,13,14,15]. One of their attributes is Degree of Similarity with Top Neighbors (DegSin) which calculates the similarity between users by using Pearson Correlation Coefficient.…”