A large language model-based detection method for poisoning attacks in recommender systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Recommender systems face poisoning attacks from malicious users injecting fake profiles.
• Current detection methods often overlook item label semantics, limiting their effectiveness against sophisticated attacks.

Purpose of the Study:

• To develop an advanced poisoning attack detection method that integrates item label semantics with user rating behaviors.
• To enhance the accuracy and robustness of identifying malicious users in recommender systems.

Main Methods:

• Utilized a pretrained large language model (LLM) for semantic encoding of item labels.
• Fused LLM-encoded label semantics with rating data into a 3D user-item-label tensor.
• Implemented a local-global joint feature extraction framework with 3D depthwise separable convolution and multi-head Performer.
• Incorporated contrastive learning to improve the separability of representations for malicious users.

Main Results:

• PAD-LLM demonstrated superior detection performance compared to baseline methods on MovieLens-1M and Amazon datasets.
• The method effectively identifies malicious users by jointly analyzing semantic and rating information.
• The proposed approach shows improved robustness under various poisoning attack scenarios.

Conclusions:

• PAD-LLM offers a significant advancement in defending recommender systems against poisoning attacks.
• Integrating LLM-based label semantics is crucial for effective detection of complex attacks.
• The method provides a robust framework for unified modeling of heterogeneous recommender system data.