Federated Learning for Privacy-Preserving Personalized Advertising Recommendation

Abstract

 This paper addresses the problem of personalized advertising recommendation under privacy constraints and proposes an optimization method based on federated learning to mitigate privacy leakage and compliance risks in centralized modeling. A global–local collaborative framework is constructed in a distributed environment, where training is performed locally on user devices and only model parameters are uploaded instead of raw data, thus enabling effective cross-device data utilization while preserving privacy. The method introduces a decoupled modeling strategy that combines global shared representations with user-specific parameters to balance individual differences and overall generalization. Regularization constraints, gradient compression, and sparsification mechanisms are adopted to improve communication efficiency and ensure stable convergence. Sensitivity analyses are conducted on multiple factors, including regularization coefficient, hidden layer dimension, data imbalance ratio, and number of communication rounds. The results show that the method achieves strong performance on Precision, F1-Score, Hit-Rate, and NDCG, and maintains stability and robustness under different uncertainty conditions. The study demonstrates the compatibility of privacy protection and personalized modeling, offering a feasible path for data utilization in advertising recommendation scenarios.