Ruan_cmu_0041E_10797.pdf (3.59 MB)
Download file

Enabling Collaborative Use of Data for Mobile Devices under Physical and Privacy Constraints

Download (3.59 MB)
thesis
posted on 17.11.2022, 21:57 authored by Yichen RuanYichen Ruan

In the mobile Internet era, data has become the essential ingredient for numerous mobile services such as video streaming, GPS navigation, and on-device intelligence. Nevertheless, the resource-restricted nature of mobile devices often obstructs them from collecting and processing all the data as needed individually. Enabling the collaborative use of data among mobile devices hence greatly extends the potential reach of mobile users for external datasets and computation resources, allowing both public and private data to be more efficiently delivered to and processed by various mobile applications. Implementing this collaboration in the real-world however faces both the physical capacity limits and the privacy protection requirements. Fortunately, the state-of-the-art edge computing framework empowered by 5G-based heterogeneous networks and the recent advances in privacy-preserving federated learning algorithms provide powerful tools based on which innovative systems and algorithms can be developed to address these constraints.

This thesis aims to enable the collaborative use of data for mobile devices through five approaches that fall into two categories: the direct sharing of public or desensitized data, and variants of privacy-preserving federated learning algorithms that are optimized for the mobile environment. For the first category, this thesis proposes to empower the direct data sharing by 1) temporal- and spatial-adaptive mobile caching and 2) topology-aware device-to-device data offloading. For the second category, this thesis improves the existing federated learning implementations by 3) the optimal client recruitment that balances both accuracy and efficiency metrics, 4) innovative extensions to federated training algorithms that incorporate flexible device participation patterns, and 5) soft clustered federated learning that also learns personalized models for clients that collect samples from multiple data distributions.

History

Date

18/03/2022

Degree Type

Dissertation

Department

Electrical and Computer Engineering

Degree Name

  • Doctor of Philosophy (PhD)

Advisor(s)

Carlee Joe-Wong

Usage metrics

Exports