Advancing Secure and Efficient Distributed Machine Learning for Edge Systems

Aug 2, 2026

This dissertation advances secure and efficient distributed machine learning for edge computing environments. It addresses challenges of deploying hybrid edge-cloud ML systems while meeting privacy and data protection requirements, including threat models and regulatory constraints. The work introduces Croesus, a multi-stage processing and transaction framework for video analytics that balances accuracy and performance. It also presents HUSH, a shared-layer approach for sharding-based machine unlearning with leakage analysis and differentially private mechanisms.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages128

Reading Time5 h 23 min

AccessFree to download

Introduction

Research Gaps and Objectives

Contributions

Background

Edge Computing and Hybrid Edge-Cloud Systems

Distributed Machine Learning

Privacy and Data Protection Regulations

Machine Unlearning

Croesus: Multi-Stage Processing and Transactions for Video-Analytics in Edge-Cloud Systems

System and Programming Model

Accuracy-Performance Trade-off in Object Detection

Croesus Design

Multi-Stage Transactions

Croesus Evaluation

Conclusion

HUSH: The Case for Shared Layers in Sharding-Based Machine Unlearning

Overview

HUSH Design

Information Leakage Analysis

Evaluation

Related Work

Conclusions and Future Directions

UC Santa Cruz UC Santa Cruz Electronic Theses and Dissertations Title Advancing Secure and Efficient Distributed Machine Learning for Edge Systems Permalink [https://escholarship.org/uc/item/6468c70q](https://escholarship.org/uc/item/6468c70q) Author Gazzaz, Samaa Publication Date 2024 Peer reviewed|Thesis/dissertation [eScholarship.org](eScholarship.org) Powered by the California Digital Library University of California UNIVERSITY OF CALIFORNIA SANTA CRUZ ADVANCING SECURE AND EFFICIENT DISTRIBUTED MACHINE LEARNING FOR EDGE SYSTEMS A dissertation submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in COMPUTER SCIENCE AND ENGINEERING by Samaa Tariq Gazzaz December 2024 The Dissertation of Samaa Tariq Gazzaz is approved: | Katia Obraczka, Chair | | --- | | Faisal Nawab | Alvaro Cardenas Peter Biehl Vice Provost and Dean of Graduate Studies Copyright © by Samaa Tariq Gazzaz 2024 Contents Figures and Tables v Abstract vii Dedication ix Acknowledgments x 1 Introduction 1 1. 1 Introducion and Motivation . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Research Gaps and Objectives . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2 Background 8 2.1 Edge Computing and Hybrid Edge-Cloud Systems ............ 8 2.1.1 Challenges in Edge-Cloud ML Systems ............. 9 2.2 Distributed Machine Learning ...................... 10 2.3 Privacy and Data Protection Regulations ................. 12 2.3.1 Overview of Data Privacy Regulations .............. 13 2.3.2 Privacy Attacks .......................... 13 2.4 Machine Unlearning ........................... 14 2.4.1 Overview of Machine Unlearning ................ 14 2.4.2 Privacy-Preserving Approaches ................. 15 2.4.3 Sharding-Based Machine Unlearning .............. 15 3 Croesus: Multi-Stage Processing and Transactions for Video-Analyticsin Edge-Cloud Systems 17 3.1 System and Programming Model ..................... 20 3.2 Accuracy-Performance Trade-off in Object Detection .......... 24 3.3 Croesus Design .............................. 25 3.3. 1 Initial-Final subsection Interaction . . . . . . . . . . . . . . . . 26 3.3.2 Algorithms ............................ 27 3.3.3 Bandwidth Thresholding ..................... 31 Contents 3.3.4 Generalizing Multi-Stage Processing ............... 34 3.4 Multi-Stage Transactions ......................... 35 3.4.1 Multi-Stage Transaction Model ................. 35 3.4.2 Safety ............................... 35 3.4.3 Multi-Stage Serializability (MS-SR) ............... 37 3.4.4 Multi-Stage Invariant Confluence with Apologies (MS-IA) ... 41 3.4.5 Multi-Partition Operations .................... 46 3.5 Croesus Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.5.1 Experimental setup ........................ 47 3.5.2 Experimental results ....................... 52 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4 HUSH: The Case for Shared LayersinSharding-Based Machine Unlearning 62 4. 1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.2 HUSH Design ............................... 71 4.2.1 System Model .......................... 71 4.2.2 Model Architecture Overview . . . . . . . . . . . . . . . . . . 74 4.2.3 Detailed Design .......................... 75 4.3 information leakage analysis ....................... 79 4.3.1 Information Leakage Overview ................. 79 4.3.2 Experimental Verification using MIA .............. 81 4.3.3 HUSH-DP: Differentially-Private Trunk ............. 87 4.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 4.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . 90 4.4.2 Baselines and metrics . . . . . . . . . .

FAQs

What problem does the dissertation target in edge-based distributed machine learning?

It targets secure and efficient distributed machine learning deployment in edge systems, where hybrid edge-cloud coordination and privacy protection are major constraints.

What is Croesus and what does it optimize?

Croesus is a multi-stage processing and transaction framework for video analytics in edge-cloud systems. It focuses on balancing accuracy and performance while enabling safe multi-stage behavior.

How does the dissertation handle machine unlearning securely?

It introduces HUSH for sharding-based machine unlearning using shared layers, including an information leakage analysis and a differentially private design to reduce privacy risk.

Aug 2, 2026

Uploaded by River Wang