Fairness Attacks on Recommender Systems

2026-06-27

AI Summary

This work proposes a structure-aware reinforcement learning attack against recommendation systems that have been trained with fairness constraints. The method generates synthetic user–item interactions that jointly preserve graph structural properties and sequential dependencies, while simultaneously optimizing both item preferences and sensitive attributes—such as gender—of fake users to deliberately undermine the system’s fairness mechanisms. By integrating a graph encoder, recurrent networks, attention mechanisms, and a reinforcement learning policy, this approach is the first to incorporate structure-aware modeling and sequence generation into fairness attacks. Experiments on four state-of-the-art recommender models and two real-world datasets demonstrate that the proposed attack significantly exacerbates recommendation unfairness with respect to sensitive attributes like gender.

Abstract

The unfairness of recommender systems has become a topic of concern due to its significant social and ethical implications. Although existing works have shown the effectiveness of attacks on the performance of recommender systems (e.g., promotion and demotion attack), the study of fairness attacks on recommender systems remains largely under-explored. To this end, we propose a novel structure-aware reinforcement learning-based fairness attack method designed to exacerbate the unfairness of target recommender systems. Specifically, we first employ a graph-based structure encoder to model the structural dependencies among the generated fake user-item interactions and the original user-item interactions. Then, we model the sequential dependency of the injected fake items using a recurrent neural network. Based on the learned structure-aware and sequence-aware representations of the fake user and item, the item selection policy attentively decides the next injected fake item. Since the target recommender system may employ fairness-aware training and leverage the user's sensitive attribute information, such as gender, we further designed a gender selection policy to decide the gender of the entire fake user profile. Both the item selection and gender selection policy are learned jointly in our proposed method. Finally, experimental results on four types of target recommendation models and two real-world datasets demonstrate the effectiveness of the proposed attack method in exacerbating the unfairness of recommender systems.

Problem

• fairness attacks
• recommender systems
• unfairness
• sensitive attributes
• adversarial attacks

Innovation

• fairness attack
• structure-aware reinforcement learning
• graph-based encoder
• sensitive attributes
• adversarial recommendation