Abstract
PrivGNN is a novel secure two-party computation (2PC) framework optimized for privacy-preserving inference on Graph Neural Networks (GNNs). By integrating additive and function secret sharing within an offline-online paradigm, PrivGNN significantly mitigates the communication and computational overhead typically associated with non-linear operations in secure computation. Experimental results demonstrate that PrivGNN achieves up to 4.7x faster online inference compared to state-of-the-art methods while maintaining high accuracy across various graph and image datasets.
Introduction
Graph Neural Networks (GNNs) have become essential for analyzing graph-structured data, ranging from molecular structures in drug discovery to social network connections. However, the sensitivity of this data often prevents organizations from leveraging cloud-based Machine Learning as a Service (MLaaS). Existing Secure Inference (SI) protocols, primarily designed for Convolutional Neural Networks (CNNs), struggle with GNNs due to structural leakage and the high computational cost of non-linear activations. PrivGNN addresses these challenges, delivering secure cloud-based graph inference up to 4.7x faster than existing methods with negligible accuracy loss.
Challenges in Secure GNN Inference
Applying standard cryptographic primitives to GNNs presents two major hurdles:
- Structural Leakage: Many existing methods fail to fully protect the relational structure of the graph, potentially leaking metadata such as the maximum node degree.
- Computational Bottlenecks: GNNs rely heavily on non-linear activation functions. Implementing these via Homomorphic Encryption or Garbled Circuits is computationally intensive and requires significant communication rounds, making them unscalable for large-scale graphs.
The PrivGNN Framework
PrivGNN operates on a client-server model under a semi-honest assumption, where parties follow the protocol but may attempt to learn private information. The framework utilizes an offline-online paradigm to push heavy cryptographic computations to a data-independent offline phase, ensuring a fast online phase.
Key Technical Innovations
- Hybrid Secret Sharing: PrivGNN combines Additive Secret Sharing (AddSS) for linear operations with Function Secret Sharing (FuncSS) for non-linear operations.
- Secure Matrix Multiplication (SMatMul): This protocol is optimized to reduce auxiliary randomness, streamlining the offline generation of Beaver triples.
- Secure DReLU (SDReLU): A novel protocol for ReLU and MaxPool layers that uses FuncSS to evaluate comparisons (e.g., $x > 0$) in a single round of online communication, bypassing the overhead of Garbled Circuits.
- Secure Piecewise Polynomials (SPiePol): To handle Sigmoid and Tanh activations, PrivGNN approximates these functions using 12-piece quadratic polynomials, which are evaluated securely via SDReLU indicators.
Experimental Evaluation
The authors benchmarked PrivGNN using standard image datasets (MNIST, CIFAR-10/100) for CNN architectures and the QM9 dataset (134,000 organic molecules) for Message Passing Neural Networks (MPNNs).
Key Results
- Performance Gains: PrivGNN achieves 1.5x to 1.7x speedups for linear layers and 2x to 15x speedups for non-linear layers compared to state-of-the-art solutions like Delphi and CrypTFlow2. Overall, online secure predictions are 1.3x to 4.7x faster.
- High Accuracy: Secure approximations have minimal impact. On the QM9 task, PrivGNN achieved 96.42% accuracy compared to 97.93% in plaintext. On MNIST, the accuracy drop was only 0.02%.
- Reduced Communication: The online communication cost for the
SQuaPolprotocol is 2x to 4x lower than competing frameworks.
Limitations and Future Work
While PrivGNN offers significant online speedups, it involves a minor accuracy trade-off due to fixed-point arithmetic truncation and polynomial approximations. Additionally, the offline phase remains communication-intensive. Future research may focus on further optimizing the offline phase, extending the framework to handle malicious adversaries, and scaling to even larger, more complex graph datasets.
Summary updated