A Privacy-Preserving Federated Recommender System with Neuro-Fuzzy Modeling and Local Differential Privacy

Abstract

The increasing reliance on personalized recommender systems in e-commerce platforms has brought major challenges in terms of data privacy, interpretability, and the model’s sturdiness. Traditional recommender systems are often mandated to access raw user data to predict relevant preferences, leading to severe privacy risks. Federated recommender system is a promising paradigm that enables decentralized model training to rectify the limitations by not exposing the unmodified user data, yet faces central server’s inference attacks and lacks in defending predictions. To bridge these research gaps, this study attempts a novel hybrid neuro–fuzzy model that integrates the strength of fuzzy membership function with a deep neural network in the federated learning environment to enhance transparency in personalized recommendations. Additionally, local differential privacy is employed with Laplace noise injection to the locally trained model gradients, thereby maintaining user privacy without revealing sensitive information. The model has been collaboratively trained on local client devices, aligning with the concepts of decentralized learning. Extensive experiments were undertaken on the real-world MovieLens 100K and 1M datasets to examine the efficacy of the presented mechanism. Research findings highlight that the studied neuro-fuzzy architecture surpasses the conventional models in terms of Normalized discounted cumulative gain, precision, recall, root mean squared error and mean absolute error over multiple measurements of privacy budget. The proposed approach achieved a strong balance among relevancy accuracy (0.205 obtained for the 100 K dataset and 0.165 achieved for the 1 M dataset), privacy and interpretability.