Data Poisoning Attacks to Session-Based Recommender Systems

Recommender systems are used to assist users in discovering their interests from websites. Companies deployed various types of recommender systems, including content-based, collaborative filtering-based, and session-based. Especially, session-based recommender systems have been deployed successfully in industry. In this work, we conduct the systematic study on data poisoning attacks to session-based recommender systems. An attacker's goal is to promote a target item such that it can be recommended to as many people as possible. Our attack injects fake users with carefully crafted interaction sessions (e.g., clicking sessions) into the recommender system to achieve this goal. The critical challenge is to choose and arrange the items in interaction sessions. We formulate our attack as an optimization problem to address this challenge, so that the injected sessions would maximize the number of users to whom the target items are recommended. We evaluate our experiments on several real-world datasets, which show that our attack methods outperform existing methods.