Penalized GANs with latent perturbation for robust shilling attack generation in recommender systems

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Shilling attacks threaten recommender system integrity by injecting fake user profiles.
• Existing generative methods struggle with unstable training and unrealistic profile generation.

Purpose of the Study:

• To propose PGAN, a Penalized Generative Adversarial Network, for generating high-quality, diverse, and undetectable shilling attack profiles.
• To enhance the robustness and realism of generated attack profiles using latent space perturbations.

Main Methods:

• PGAN utilizes a gradient penalty to stabilize discriminator training.
• Controlled noise perturbations are applied in the generator's latent space for improved robustness and diversity.
• Evaluation on real-world datasets using metrics like Hit Ratio@K and Prediction Shift.

Main Results:

• PGAN consistently outperforms traditional statistical attacks and baseline GAN models.
• Achieved HR@10 scores of 0.2051 on MovieLens and 0.2076 on Amazon datasets.
• Generated profiles demonstrate high similarity to genuine users, confirming their realism.

Conclusions:

• PGAN offers a superior approach for generating realistic and diverse shilling attack profiles.
• The proposed method enhances recommender system security by providing a robust attack generation framework.
• PGAN represents a significant advancement over existing generative adversarial network-based attack methods.