[Paper Review] A Vertical Federated Learning Framework for Graph Convolutional Network

Xiang Ni, Xiaolong Xu+3 more|arXiv (Cornell University)|Jun 22, 2021

Privacy-Preserving Technologies in DataComputer Science23 references26 citations

TL;DR

FedVGCN enables privacy-preserving vertical federated learning for graph convolutional networks using additively homomorphic encryption, achieving close to centralized performance for GraphSAGE across benchmarks.

ABSTRACT

Recently, Graph Neural Network (GNN) has achieved remarkable success in various real-world problems on graph data. However in most industries, data exists in the form of isolated islands and the data privacy and security is also an important issue. In this paper, we propose FedVGCN, a federated GCN learning paradigm for privacy-preserving node classification task under data vertically partitioned setting, which can be generalized to existing GCN models. Specifically, we split the computation graph data into two parts. For each iteration of the training process, the two parties transfer intermediate results to each other under homomorphic encryption. We conduct experiments on benchmark data and the results demonstrate the effectiveness of FedVGCN in the case of GraphSage.

Motivation & Objective

• Motivate privacy-aware learning when graph data are vertically partitioned across entities.
• Propose FedVGCN framework to train GCN-based models without sharing raw features or edges.
• Show that additively homomorphic encryption preserves accuracy while protecting data privacy.
• Demonstrate that FedVGCN outperforms isolated-data baselines and approaches centralized plaintext performance.

Proposed method

• Split graph data into two feature-holding parties in a vertical FL setting.
• Use additively homomorphic encryption (Paillier) to exchange encrypted intermediate results during forward and backward passes.
• Approximate ReLU with quadratic orthogonal polynomials to preserve homomorphic addition and multiplication.
• Apply a second-order Taylor approximation for loss and gradient computations to enable encrypted operations.
• Leverage Private Set Intersection (PSI) for initial ID alignment before training.
• Evaluate on GraphSAGE extensions with unsupervised loss inspired by graph-based negative sampling.

Experimental results

Research questions

• RQ1 Can FedVGCN improve node classification accuracy when data are vertically partitioned compared to isolated-training graphs?
• RQ2 How close can FedVGCN get to a centralized model trained on combined plaintext data?
• RQ3 What is the impact of using Paillier HE and polynomial activations on privacy and accuracy in vertical GNN training?

Key findings

• FedVGraphSage consistently outperforms GraphSage models trained on isolated feature/edge data across Cora, PubMed, and Citeseer.
• FedVGraphSage achieves accuracy close to GraphSage trained on combined plaintext data (e.g., 0.6770 vs 0.7080 on Cora; 0.7830 vs 0.7890 on PubMed; 0.6820 vs 0.6983 on Citeseer).
• Additively homomorphic encryption enables privacy-preserving communication of intermediate results without revealing raw data.
• Quadratic polynomial activations enable ciphertext-friendly computation, preserving sums and multiplications under encryption.
• The framework supports privacy guarantees under the assumption of non-colluding server and clients with security grounded in established scalar-product protocols.