Influence-Driven Data Poisoning for Robust Recommender Systems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommender systems are vulnerable to data poisoning attacks, leading to biased recommendations.
• Assessing the threat of bi-level poisoning and ensuring user imperceptibility are key challenges.
• Existing attack methods struggle with efficient threat assessment and generating unnoticeable fake data.

Purpose of the Study:

• To propose an efficient poisoning attack strategy, Infmix, that addresses threat assessment and user imperceptibility.
• To introduce a novel defense strategy, adversarial poisoning training (APT), to improve recommendation system robustness.

Main Methods:

• Infmix utilizes an influence-based threat estimator and a user generator (Usermix) for efficient attack execution.
• Usermix generates unnoticeable fake data, even with limited known users.
• Adversarial poisoning training (APT) mimics poisoning by injecting fake users to minimize empirical risk, using influence functions for optimization.

Main Results:

• Infmix demonstrates superiority in attacking six recommendation systems across four real datasets.
• Extensive experiments validate the effectiveness of Infmix in executing efficient and imperceptible poisoning attacks.
• Adversarial poisoning training (APT) significantly improves recommendation robustness against poisoning attacks.

Conclusions:

• Infmix provides an effective solution for bi-level poisoning attacks, overcoming limitations of existing methods.
• Adversarial poisoning training (APT) offers a robust defense mechanism against sophisticated recommendation system attacks.
• The proposed methods enhance the security and reliability of recommender systems.