Graduation Year

2022

Document Type

Thesis

Degree

M.A.

Degree Name

Master of Arts (M.A.)

Degree Granting Department

Mathematics and Statistics

Major Professor

Kaiqi Xiong, Ph.D.

Committee Member

Lu Lu, Ph.D.

Committee Member

Razvan Teodorescu, Ph.D.

Committee Member

Jiwoong Kim, Ph.D.

Keywords

Data Augmentation, Federated Learning, Machine Learning/Deep Learning, Neural Tangent Generalization Attack

Abstract

Nowadays, a massive amount of data is generated and stored on servers and cloudsfrom various applications daily. Preventing these data from unauthorized use often becomes necessary and critical in various real-world applications. Many researchers have studied this crucial problem and developed different methods for this purpose. Among them, Neural Tangent Generalization Attack (NTGA) is one of the most efficient methods to make a dataset unlearnable, which means that the dataset is not learnable by machine learning/deep learning methods. That is, the NTGA-generated dataset is protected against unauthorized use. In this thesis, we explore the vulnerability of an NTGA-generated unlearnable CIFAR-10 dataset. Specifically, we show that by employing data augmentation, we can still train a model using the CIFAR-10 dataset generated by NTGA in a centralized learning environment. That is, we experimentally prove that the NTGA-generated unlearnable CIFAR-10 dataset is vulnerable or learnable. To reduce the execution time of learning from the dataset, we further study the vulnerability of the unlearnable dataset in a decentralized learning (i.e., federated learning) environment. Our experiments demonstrate the efficiency of the proposed decentralized method. After all, we reduce the execution time for learning from the unlearnable dataset when using decentralized learning compared to centralized learning.

Download

Share

COinS