CF-DPGNN: An Edge-Level Differential Privacy Framework for Collaborative Filtering Recommendation System

Recently, Graph Neural Network (GNN)-based collaborative filtering recommendation systems (CFRSs) have faced severe privacy leakage risks. However, existing differential privacy (DP) methods primarily rely on the traditional differentially private stochastic gradient descent (DPSGD) framework, which causes noise explosion during training and severely degrades recommendation accuracy. To address this issue, we propose CF-DPGNN, a privacy amplification-based DP protection framework for collaborative filtering GNNs. CFDPGNN integrates subgraph sampling with DPSGD to realize privacy amplification mechanism, aiming to enhance the privacyutility trade-off and protect sensitive information in recommendation systems (RSs). Experimental results on three real-world online platform datasets demonstrate that CF-DPGNN achieves SOTA, outperforming the second-best method by 9.7% and 10.9% in key metrics, while improving privacy attack defense capabilities by 10.2% and 11.6%. These findings validate the effectiveness of CF-DPGNN.