LPGNN: Implementing Local Differential Privacy Protection in Graph Neural Networks

LPGNN is the first research project to systematically address the Local Differential Privacy (LDP) problem in Graph Neural Networks (GNNs). Developed by Sina Sajadmanesh, its related paper was published at the ACM CCS 2021 conference. This project implements a complete framework that encrypts and protects user data before it leaves the device while maintaining high model accuracy. The code is open-sourced on GitHub (https://github.com/sisaman/LPGNN), with the last update date being June 10, 2026.

Graph data is widely present in scenarios like social networks and e-commerce interactions, containing sensitive personal information. Traditional privacy protection relies on trusted centers, which are prone to data leaks; LDP, on the other hand, allows local perturbation of data, so servers only receive noisy data. However, node features in GNNs are closely related to their neighbors, and traditional LDP mechanisms are not suitable for this structural dependency, leading to complexity in privacy budget allocation and noise injection.

Core Insight of LPGNN: Node representations are the result of aggregating their own and neighbors' features, so noise must be injected when reporting features and its propagation controlled. Key Components: 1. Feature-level LDP Mechanism: For continuous features, use random response and dimension partitioning, with subspaces protected independently; 2. Privacy-Aware Architecture: Redesign message passing, with adaptive aggregation weights and inter-layer noise calibration. The technical implementation supports architectures like GCN and GAT, with local perturbation requiring no central trust and fine-grained management of privacy budgets.

Evaluated on datasets including Cora, Citeseer, Pubmed (citation networks), and Facebook Page-Page (social networks): When ε=8, the accuracy of LPGNN on Cora differs by less than 5% from the non-private model, outperforming the baseline with direct noise addition. Moderate noise may also act as regularization to improve generalization ability.

LPGNN provides a foundation for federated graph learning and can be used in scenarios like financial risk control (joint fraud detection) and medical diagnosis (joint disease prediction). It complies with regulations such as GDPR and CCPA, helping enterprises turn privacy protection from a compliance cost into a competitive advantage.

Current Limitations: Only supports homogeneous graphs; has not explored heterogeneous graphs (e.g., knowledge graphs) or dynamic graphs; in terms of computational efficiency, the local perturbation overhead for ultra-large-scale graphs is high. Future Directions: Expand to heterogeneous/dynamic graphs, design efficient perturbation algorithms or hardware acceleration.

LPGNN promotes the evolution of privacy-preserving machine learning from centralized to local protection, proving that GNNs can still maintain learning capabilities under strict local privacy. For researchers and engineers, it is both a tool and a conceptual framework that will unlock the potential of graph intelligence while protecting privacy.