Advancing Recommender Systems by mitigating Shilling attacks

TL;DR

The paper addresses the vulnerability of recommender systems to shilling attacks that bias predictions. It introduces an unsupervised PCA-based detection method that identifies shilling profiles by leveraging high inter-profile covariance, using correlation and profile thresholds to flag attackers. Experiments on MovieLens 100K across multiple attack models demonstrate that detection can be highly accurate when attacker presence crosses a defined threshold, with user-based similarity showing greater susceptibility to attack than item-based approaches. The approach offers a robust, model-agnostic defense that can be integrated with various recommender systems to preserve recommendation integrity and user trust.

Abstract

Considering the premise that the number of products offered grow in an exponential fashion and the amount of data that a user can assimilate before making a decision is relatively small, recommender systems help in categorizing content according to user preferences. Collaborative filtering is a widely used method for computing recommendations due to its good performance. But, this method makes the system vulnerable to attacks which try to bias the recommendations. These attacks, known as 'shilling attacks' are performed to push an item or nuke an item in the system. This paper proposes an algorithm to detect such shilling profiles in the system accurately and also study the effects of such profiles on the recommendations.

Figures (15)

• Figure 1: Similarity between users
• Figure 2: Similarity between items
• Figure 3: Difference between collaborative and content based filtering
• Figure 4: Model Based Filtering
• Figure 5: The illustration of a shilling profile