Robust Privacy-Preserving Recommendation Systems Driven by Multimodal Federated Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Privacy

Background:

• Recommendation systems (RS) are crucial but face privacy challenges.
• Federated learning (FL) offers a privacy-preserving approach for RS.
• Existing FL-based RS often lack multimodal data integration and robust security against malicious clients.

Purpose of the Study:

• To design a federated learning-based, privacy-preserving, multimodal recommendation system framework.
• To address the limitations of unimodal learning and enhance recommendation accuracy using diverse data modalities.
• To develop robust defenses against malicious clients and Byzantine attacks in federated recommendation.

Main Methods:

• An attention mechanism to effectively integrate and weight features from multiple data modalities.
• Local differential privacy (LDP) for enhanced data protection at the client level.
• Personalized federated learning strategies to identify and mitigate malicious client behavior and Byzantine attacks.
• Development and validation using two novel multimodal datasets.

Main Results:

• The proposed framework demonstrates superior performance in recommendation accuracy compared to existing methods.
• Effective integration of multimodal data significantly boosts recommendation quality.
• The implemented privacy-preserving techniques successfully defend against data inference and Byzantine attacks.
• Simulation results confirm the robustness and effectiveness of the designed FL-based multimodal RS.

Conclusions:

• The developed FL-based multimodal recommendation system offers a significant advancement in privacy-preserving AI.
• The integration of attention mechanisms and advanced privacy strategies provides a robust solution for secure and accurate recommendations.
• This work paves the way for more sophisticated and trustworthy recommendation systems in sensitive data environments.