Shilling Black-Box Recommender Systems by Learning to Generate Fake User Profiles

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommender systems (RS) are crucial for guiding customer purchases.
• Unscrupulous parties may exploit RS through shilling attacks using fake user profiles.
• Existing shilling attack methods often lack transferability or invisibility.

Purpose of the Study:

• To develop a novel attack model for generating undetectable and transferable fake user profiles.
• To overcome the limitations of conventional shilling attack approaches.
• To improve the effectiveness of shilling attacks against recommender systems.

Main Methods:

• Introduced Leg-UP, a generative adversarial network (GAN)-based model for fake profile generation.
• Leg-UP learns user behavior from real user data (templates) to construct realistic fake profiles.
• Employed a generator to output discrete ratings and a discriminator for invisibility, optimizing for transferability on a surrogate RS model.

Main Results:

• Leg-UP successfully generates discrete ratings to simulate real users.
• The model demonstrates enhanced attack transferability by optimizing generator parameters on a surrogate RS model.
• Experiments show Leg-UP surpasses state-of-the-art shilling attack methods across various RS models.

Conclusions:

• Leg-UP offers a significant advancement in shilling attack methodologies for recommender systems.
• The proposed model effectively balances attack transferability and invisibility.
• Leg-UP provides a robust solution for generating sophisticated fake user profiles in RS security research.