Coarsening to Conceal: Enabling Privacy-Preserving Federated Learning for Graph Data

Ekta Srivastava, Manoj Kumar, Kumar Arjun, Sandeep Kumar

Keywords: Federated Learning, Privacy-Preserving Machine Learning, Graph Neural Networks, Graph Coarsening, Data Privacy and Security

TL;DR: We propose a novel approach to enhancing privacy in federated learning by leveraging graph coarsening, demonstrating its effectiveness in concealing sensitive information while maintaining robust collaborative learning across decentralized entities.

Abstract: With the escalating demand for privacy-preserving machine learning, federated learning (FL) stands out by enabling collaboration among decentralized entities. Utilizing graph representations of data enhances learning for graph-level tasks, crucial for FL with data distributed across local repositories. Despite its benefits, stringent privacy regulations often compromise FL's performance. Previous methods aimed at ensuring privacy introduce performance degradation and computational overhead. In response to these challenges, we propose using graph coarsening—a simple yet effective method—to enhance the security and privacy of FL on graph data. Our approach posits that graph coarsening alone can suffice for privacy guarantees, as model parameters obtained from training on the coarsened graph effectively conceal sensitive information susceptible to privacy attacks. Through comprehensive application and analysis, we demonstrate the efficacy of graph coarsening within an FL setup, taking both the graph matrix and node features as input, and jointly learning the coarsened graph matrix and feature matrix while ensuring desired properties. The resultant coarsened graph representations are then utilized to train model parameters, subsequently communicated within an FL framework for downstream tasks such as classification. Extensive experimentation across various datasets confirms that graph coarsening ensures privacy while enhancing performance with minimal trade-offs compared to traditional differential privacy (DP) methods without adding extra complexity overhead.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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: 11943