Differentially Private Graph Convolutional Networks with Privacy Amplification

This study delves into the application of Differential Privacy (DP) in Graph Convolutional Networks (GCN). We initially discussed the limitations of existing approaches to privacy-preserving graph learning. These limitations primarily stem from an unsatisfied privacy-utility trade-off, as well as their constrained performance across diverse datasets. To address these challenges, we propose a novel approach that incorporates DP with privacy amplification techniques and combines it with Renyi Differential Privacy (RDP) and differentially private stochastic gradient descent (DPSGD). Our approach aims to enhance the privacy-utility trade-off and demonstrate universality across various types of datasets. Through carefully designed privacy attack experiments and comparative baseline evaluations, we demonstrate that compared to existing methods, ours exhibits superior utility with stronger privacy protection in multiple datasets, and sampling-based privacy amplification techniques contribute to improved model expressiveness. In summary, our work contributes to a general and versatile approach to achieving privacy-preserving GCN, offering a better privacy-utility trade-off compared to SOTA.