Toward Secure Graph Data Collaboration in a Data-Sharing-Free Manner: A Novel Privacy-Preserving Graph Pre-training Model
Preprint
- 1 January 2023
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Graph data is a valuable data source that has been widely adopted in a variety of domains, such as bioinformatics, social networks, and finance. Such data is unique in that it increases in value through data sharing. However, sharing raw graph data can pose risks to information security and commercial confidentiality. To mitigate such risks, it is more desirable to share graph pre-trained models that contain knowledge derived from the data. Recent studies have shown that even graph pre-trained models come with some privacy risks, as an adversary with access to the model can reveal private information in the training graph. To address this issue and move towards a safer graph knowledge-sharing environment, we propose a privacy-preserving graph pre-training model for sharing graph information. In particular, we introduce a novel principle of privacy-preserving data augmentation, which can be paired with graph contrastive learning for pre-training a privacy-preserving graph neural network. The privacy-preserving data augmentation enables graph models to obfuscate private information during training. Extensive experiments suggest that our proposed model outperforms state-of-the-art approaches in reducing privacy risk while maintaining generalizability. Our further case studies offer additional insights into how graph properties affect generalizability and privacy-preserving performance.Keywords
This publication has 25 references indexed in Scilit:
- Multi-Scale attributed node embeddingJournal of Complex Networks, 2021
- Pre-trained models: Past, present and futureAI Open, 2021
- Random Walks: A Review of Algorithms and ApplicationsIEEE Transactions on Emerging Topics in Computational Intelligence, 2019
- Characterizing user navigation and interactions in online social networksInformation Sciences, 2012
- Identifying influential nodes in complex networksPhysica A: Statistical Mechanics and its Applications, 2012
- Link prediction in complex networks: A surveyPhysica A: Statistical Mechanics and its Applications, 2011
- Networks2010
- Fast Random Walk with Restart and Its ApplicationsIEEE International Conference on Data Mining (ICDM), 2006
- Structural Holes and Good IdeasAmerican Journal of Sociology, 2004
- Power and Centrality: A Family of MeasuresAmerican Journal of Sociology, 1987