ERI: An Edge-Based Staleness-Resilient Data Synthesis Framework for Edge AI application in Network Measurement

As digital infrastructure grows increasingly dependent on real-time data from distributed sensors and devices, the need to monitor network traffic locally and reliably becomes crucial. This project addresses a critical challenge in edge computing, i.e., how to provide accurate and up-to-date data to Edge AI applications while preserving user privacy and minimizing communication overhead. Traditional methods struggle with outdated data, privacy limitations, and limited processing power at the edge. This research seeks to develop a novel data synthesis framework that enables edge devices to generate privacy-preserving synthetic traffic data while adapting to changes in user behavior and network conditions. The framework aims to improve real-time performance and reliability in edge applications like anomaly detection and traffic classification. Broader impacts include improved security and efficiency in edge-based systems across sectors such as healthcare, industrial control, and mobile communications. The project will also enhance undergraduate and graduate curricula in AI, cybersecurity, and edge computing, and support outreach programs to engage K-12 students.

This project will create a staleness-resilient, privacy-aware data synthesis framework for Edge AI-based network measurement. It combines quality-controlled synthetic data generation with decentralized collaboration among edge devices. The first research thrust focuses on quantifying and optimizing the trade-off between synthetic data usability and privacy via a learnable feature-discarding mechanism within an autoencoder-based system. The second thrust introduces a collaborative, low-overhead generative adversarial network architecture, where feedback from edge devices is privatized via local differential privacy mechanisms. The framework also includes an adaptive staleness detection system that uses model confidence and uncertainty analysis to refresh data generators in response to shifting data patterns. A small-scale testbed will validate system performance in realistic edge environments. By advancing methods in privacy-preserving learning, data generation, and distributed collaboration, this work establishes new foundations for using distributed AI on data synthesis in sensitive, dynamic environments.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Award Number: 2502132
Principal Investigator: Jielun Zhang
Funds Obligated: $199,993
State: ND