Date of Award
2024
Document Type
Open Access Dissertation
Degree Name
Doctor of Philosophy in Computer Engineering (PhD)
Administrative Home Department
Department of Electrical and Computer Engineering
Advisor 1
Kaichen Yang
Committee Member 1
Lan Zhang
Committee Member 2
Hongyu An
Committee Member 3
Yu Cai
Abstract
In the evolving landscape of machine learning and artificial intelligence, this dissertation presents a series of innovative contributions spanning several critical areas: embracing semi-supervised domain adaptation for secure knowledge transfer, enhancing the model performance of tiny models, and executing model stealing attacks via diversified prompts. The overarching goal is to enhance the performance, scalability, and security of AI models across various applications.
The first research focus is on semi-supervised domain adaptation within federated learning frameworks. By leveraging semi-supervised learning techniques, this work addresses the challenge of adapting models trained on a source domain to perform effectively on a target domain with differing data distributions. This approach ensures the secure transfer of knowledge across domains without compromising data privacy. The second area of research explores the augmentation of heterogeneous models, particularly focusing on the integration of tiny and large models within federated learning environments. This innovative method demonstrates significant improvements in the efficiency and effectiveness of TinyML, enabling their deployment in resource-constrained environments. Lastly, the contribution is in the realm of secure model extraction and utilization. This research delves into the vulnerabilities of fine-tuned foundation models, proposing a novel attack method, termed "Siemese attack," that uses diversified prompts to steal the functionality of these models. The study highlights the importance of securing AI models against such threats while also exploring methods to safeguard the confidentiality of task-specific knowledge embedded within these models. Overall, the research presented in this dissertation advances the state-of-the-art in federated learning, tiny model performance, model stealing and secure AI practices. The findings have significant implications for the development of robust, efficient, and secure AI systems capable of addressing diverse and complex real-world challenges.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Das, Madhureeta, "ENHANCING PRIVACY WHILE REVEALING VULNERABILITIES: STRATEGIES FOR ADAPTATION, OPTIMIZATION, AND MODEL EXTRACTION", Open Access Dissertation, Michigan Technological University, 2024.