# Multi-criteria group shilling attacks

## TL;DR

This paper introduces a new type of attack on group recommendation systems and shows how to make them more robust against such attacks.

## Contribution

The paper proposes a novel multi-criteria group shilling attack model and evaluates its impact on system robustness.

## Key findings

- The proposed multi-criteria system improves robustness with up to 12% higher average hit ratio on the YM20 dataset.
- The MUP-NNZ strategy shows consistent resistance to profile injection attacks.
- This work fills a critical research gap in multi-criteria group recommendation system robustness.

## Abstract

The rapid advancement of technology has enabled the collection of detailed, multi-dimensional user data, paving the way for multi-criteria recommendation systems that consider diverse aspects of user preferences. While traditional recommendation systems aim to satisfy individual users, group recommendation systems are designed to generate suggestions that accommodate the collective preferences of a group. However, the increasing prevalence of group interactions in digital environments has also introduced new vulnerabilities, such as group shilling attacks, where coordinated malicious users manipulate recommendation outcomes. This study conducts the first comprehensive robustness analysis of multi-criteria group recommender systems, addressing a critical research gap. A novel shilling attack strategy is proposed by adapting the group shilling model to multi-criteria settings, allowing a deeper understanding of the risks these systems face. Experimental results indicate that the proposed multi-criteria recommender system achieves notable robustness across datasets. Specifically, the average hit ratio (AvgHR) increases up to approximately 12% on the YM20 dataset and reaches around 15% on the YM10 dataset. Furthermore, among the target item selection strategies, the MUP-NNZ method consistently demonstrates superior resistance to profile injection attacks, confirming its effectiveness in maintaining recommendation accuracy under adversarial conditions.