Task and Model Agnostic Differentially Private Graph Neural Networks via Coarsening
Keywords: Graph Neural Network (GNN), Differential Privacy (DP), Graph Coarsening
Abstract: Graph Neural Networks (GNNs) have emerged as powerful tools for analyzing graph-structured data, deriving representations by aggregating information from neighboring nodes. However, this aggregation process inherently increases the risk of exposing confidential data, as a single node may influence the inference process for multiple nodes simultaneously. To mitigate this risk, researchers have explored differentially private training methods for GNN models. Existing privacy-preserving approaches, however, face significant challenges. They often incur high computational costs during training or struggle to generalize across various GNN models and task objectives. To address these limitations, we introduce Differentially Private Graph Coarsening (DPGC), a novel method that tackles two key challenges in GNN training: scalability and privacy guarantees that are independent of the downstream task or GNN model. Through comprehensive experiments on six datasets across diverse prediction tasks, we demonstrate that DPGC sets new benchmarks in graph coarsening. Our method achieves superior compression-accuracy trade-offs while maintaining robust privacy guarantees, outperforming state-of-the-art baselines in this domain.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10356
Loading