DPRO-GNN: Bridging differential privacy and advanced optimization for privacy-preserving graph learning

Abstract

Graph Neural Networks (GNNs) have demonstrated exceptional performance in modeling structured data, yet their application in sensitive domains inevitably raises privacy concerns. Existing Differentially Private GNN (DPGNN) frameworks primarily rely on Differentially Private Stochastic Gradient Descent (DP-SGD) to enforce privacy guarantees. However, DP-SGD inherits its inherent limitations, such as training instability and slow convergence, which are particularly problematic for complex graph learning tasks. Although advanced optimizers like Ranger offer a promising alternative, their naive integration into DPGNN frameworks introduces bias, specifically in the second-moment estimation, due to the additive noise required for DP. To address this challenge, we propose the Differentially Private Ranger-Optimized Graph Neural Network (DPRO-GNN) to protect users’ sensitive data when training the GNN tasks. To mitigate DP noise and capture multi-scale structure, DPRO-GNN applies hierarchical pooling to aggregate nodes into progressively coarser subgraphs, yielding robust, multi-resolution embeddings. Meanwhile, our approach introduces DP-RangerBC, a bias-corrected variant of the Ranger optimizer that mitigates the noise-induced bias in second-order moment estimation, thereby enabling more stable and efficient training under DP constraints. Furthermore, the theoretical analysis of DPRO-GNN, including its correctness and security, is also provided. Extensive experiments on real-world datasets demonstrate that DPRO-GNN achieves superior performance in terms of classification accuracy and convergence speed, compared to state-of-the-art DPGNN methods. The code of DPRO-GNN is available at the following link:https://github.com/Silbermondlel/DPRO-GNN.