Privrecnet: A Privacy-Preserving Decentralized Learning Framework for Recommendation Systems

Abstract

In the rapidly evolving data-driven world, user privacy protection is essential, particularly within machine learning applications. Our study introduces PrivRecNet, an innovative framework that merges the privacy-preserving strengths of decentralized learning with the sophisticated capabilities of transformer-based models, tailored for recommendation systems. Decentralized learning offers a decentralized alternative to traditional machine learning, enhancing both user privacy and data security. PrivRecNet employs two advanced transformer models: Bidirectional Encoder Representations from Transformers and Behavior Sequence Transformer, within a decentralized learning context. The framework's performance is evaluated using the Amazon Customer Review and MovieLens-1M datasets. Results are promising: the decentralized Bidirectional Encoder Representations from Transformers model achieves impressive accuracies of 87% and 76% on two distinct datasets. Similarly, the decentralizedBehavior Sequence Transformer model exhibits a mean absolute error of 0.8. Our research highlights the dual benefits of decentralized learning in boosting model accuracy and preserving user privacy. The findings demonstrate that PrivRecNet can significantly enhance recommendation system performance without compromising data privacy, marking a crucial advancement in developing effective, privacy-conscious machine learning solutions and contributing to the broader field of ethical and responsible artificial intelligence.