Exploring Leaked Information in Privacy-Aware Machine Learning - Thesis Abstract

Aug 2, 2026

Upload by Liam

Privacy-aware machine learning embeds privacy requirements into the design, development, and deployment of machine learning systems. Federated learning and machine unlearning support protecting user data and honoring the right to be forgotten, yet privacy leakage can still occur through model parameter changes. This thesis studies leakage in federated learning by exploiting leaked gradients and mobility prior knowledge to recover private data, then examines machine unlearning leakage and proposes reconstruction attacks including forgotten sample reconstruction and hierarchical attacks using pre-trained models, along with countermeasures via secure substitution.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages109

Reading Time4 h 35 min

Exploring Leaked Information in Privacy-Aware Machine Learning A thesis submitted in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Information Systems by Kaiyue Zhang to School of Computer Science Faculty of Engineering and Information Technology University of Technology Sydney NSW-2007, Australia April 2024 © 2024 by Kaiyue Zhang All Rights Reserved AUTHOR’S DECLARATION of Ce UtheienceTechy my,conySorofydkEnungeynle,ringtraltherInformation Treferenced orecanoknyl- , Kaiyue Zhang declare that this thesis is submitted in partial fulfilment of the requirements for the award of Doctor of Philosophy, in the School edged. In addition, I certify that all information sources and literature used are indicated in the thesis. I certify that the work in this thesis has not previously been submitted fora degree nor has it been submitted as part of the requirements for a degree at any other academic institution except as fully acknowledged within the text. This thesis is the result of a Collaborative Doctoral Research Degree program with the Southern University of Science and Technology. This research is supported by the Australian Government Research Training Program. Production Note: SIGNATURE: Signature removed prior to publication. [Kaiyue Zhang] DATE: 25th March, 2024 PLACE: Sydney, Australia ACKNOWLEDGMENTS I invPhalDuablejourguineydaPncrofee,sunwaversor Yu hgbseupenportselfl, anessdlyprsofounparindgsghts thffort toroughoguideme tyo am profoundly grateful to Professor Shui Yu and Professor Xuan Song for their ensure that I am on the right track of research, while Prof. Song has been a constant source of encouragement and trust. Being under the supervision of Prof. Yu and Prof. Song has been an incredibly fortunate experience. My sincere thanks to them for their patience and expertise, which have been pivotal in making this academic endeavour rewarding and enriching. Thanks to Dr. Zipei Fan, my senior, from whom I learned various methodologies and shared numerous experiences over these years. I extend my gratitude to my colleagues and academic collaborators, including Jiewen Deng, Mingjian Tang, Zhiyi Tian, Weiqi Wang, Du Yin, Chenhan Zhang and Jiaqi Zhang, among others. Learning and discussing with them have provided significant inspiration and encouragement throughout my academic journey. I cherish and appreciate this memory of working together. I also want to express my thanks to my friends, who, despite infrequent in-person meetings, have consistently been there for me online, offering immense courage and comfort. I would also like to express my gratitude to the University of Technology Sydney and the Southern University of Science and Technology for providing funding for my Ph.D. research. Lastly, my deepest appreciation goes to my parents and my boyfriend. While the Ph.D. journey is concluding, my love and gratitude for them will endure endlessly. LIST OF PUBLICATIONS 1. Kaiyue Zhang, Zipei Fan, Xuan Song, Shui Yu. Enhancing Machine Unlearning via Generative Substitution. IEEE Transactions on Dependable and Secure Computing. Under Review. 2. Kaiyue Zhang, Zipei Fan, Xuan Song, Shui Yu. Pre-trained Generative Models Guided Reconstruction Attack in Machine Unlearning. IEEE Transactions on Dependable and Secure Computing. Under Review. 3. Kaiyue Zhang, Weiqi Wang, Zipei Fan, Xuan Song, Shui Yu. Conditional Matching GAN Guided Reconstruction Attack in Machine Unlearning. IEEE Global Communications Conference, pp. 44-49, 2023 4. Kaiyue Zhang, Zipei Fan, Xuan Song, Shui Yu. Enhancing Trajectory Recovery From Gradients via Mobility Prior Knowledge. IEEE Internet of Things Journal, 10(6), 2022 5. Kaiyue Zhang, Xuan Song, Chenhan Zhang, Shui Yu. Challenges and Future Directions of Secure Federated Learning: A Survey. Frontiers of computer science, 16(5), 2022 ABSTRACT Privacy-aware machine learning is a significant field t

FAQs

What privacy risks remain in privacy-aware machine learning despite federated learning and machine unlearning?

Even when methods avoid exposing raw data and only disclose model parameters, privacy leakage can still occur through changes in those parameters.

How does the thesis approach information leakage in federated learning?

It analyzes leakage stemming from aggregated gradients, then uses leaked gradients together with mobility prior knowledge to reconstruct users’ private data via a proposed reconstruction attack algorithm.

What defense and counter-attack strategies does the thesis investigate?

It studies counter-attack strategies against reconstruction attacks and develops a new unlearning strategy based on secure substitution to improve privacy protection.

Aug 2, 2026

Upload by Liam