Georgia Institute of Technology
Understanding and Mitigating Privacy Vulnerabilities in Deep Learning
Abstract
dc:description.abstractAdvancements in Deep Learning (DL) have enabled leveraging large-scale datasets to train models that perform challenging tasks at a level that mimics human intelligence. In several real-world scenarios, the data used for training, the trained model, and the data used for inference can be private and distributed across multiple distrusting parties, posing a challenge for training and inference. Several privacy-preserving training and inference frameworks have been developed to address this challenge. For instance, frameworks like federated learning and split learning have been proposed to train a model collaboratively on distributed data without explicitly sharing the private data to protect training data privacy. To protect model privacy during inference, the model owners have adopted a client-server architecture to provide inference services, wherein the end-users are only allowed black-box access to the model’s predictions for their input queries. The goal of this thesis is to provide a better understanding of the privacy properties of the DL frameworks used for privacy-preserving training and inference. While these frameworks have the appearance of keeping the data and model private, the information exchanged during training/inference has the potential to compromise the privacy of the parties involved by leaking sensitive data. We aim to understand if these frameworks are truly capable of preventing the leakage of model and training data in realistic settings. In this pursuit, we discover new vulnerabilities that can be exploited to design powerful attacks that can overcome the limitations of prior works and break the illusion of privacy. Our findings highlight the limitations of these frameworks and underscore the importance of principled techniques to protect privacy. Furthermore, we leverage our improved understanding to design better defenses that can significantly deter the efficacy of an attack.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kariyappa, Sanjay
- Committee members dc:contributor.committeemember
-
- Qureshi, Moinuddin K.
- AlRegib, Ghassan
- Krishna, Tushar
- Prakash, Atul
- Suh, G. Edward
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- http://hdl.handle.net/1853/70130