Основные понятия
The author proposes the GraphPub framework to protect graph topology while maintaining data availability, achieving high model accuracy with low privacy budget.
Аннотация
The paper introduces GraphPub, a novel framework for protecting graph topology while ensuring data availability. By utilizing reverse learning and encoder-decoder mechanisms, false edges are introduced to replace real edges, maintaining model accuracy close to the original graph. Experimental results demonstrate the effectiveness of GraphPub in preserving privacy while maintaining high model accuracy under stringent privacy budgets. The study also explores degree preservation and scalability across different GNN models.
Key Points:
- Introduction of GraphPub for differential privacy protection in graphs.
- Utilization of reverse learning and encoder-decoder mechanisms for edge protection.
- Experimental validation showcasing high model accuracy with low privacy budget.
- Degree preservation analysis and scalability demonstrated across various GNN models.
Статистика
"Sufficient experiments prove that our framework achieves model accuracy close to the original graph with an extremely low privacy budget."
"Our model maintains a high accuracy when privacy protection requirement is extremely strict (the privacy budget ϵ is very small, ϵ = 1)."