Graduate Thesis Or Dissertation
 

Information-theoretic Approach to Design and Evaluate Privacy-preserving and Fair Frameworks for Continuous High-dimensional Data

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/t435gm99w

Descriptions

Attribute NameValues
Creator
Abstract
  • Deep learning is becoming the latest trend in sensitive applications, such as healthcare, criminal justice, and finance. As these new applications emerge, adversaries are circumventing them. Further, there have been concerns about the possibility of bias and discrimination in predictive applications. In order to address these issues, we propose an information-theoretic approach to design a continuous high-dimensional data deep learning framework. We call this framework Gaussian privacy protector (GPP). Our proposed framework has many advantages: (1) it reduces the problem to the optimal compression of data about a measure of utility and privacy; (2) it can prevent adversaries from private mining information from the released data while simultaneously maximizing the amount of the utility's information revealed; (3) it adapts the idea of the information bottleneck (IB) based on the problem of revealing data, which is often sensitive; (4) it considers a privacy funnel (PF) problem inspired by utility data as the central part of data to be revealed; (5) using a similar framework, we show how to achieve fairness in classification; and (6) this work illustrates the feasibility of creating a centralized platform to support this framework over distributed datasets. We utilize variational lower bounds of mutual information approximation implemented as supervised learning using an adversarial training algorithm. We use three datasets: hand-written digits (MNIST), celeb faces attributes (CelebA), and human activities and postural transitions' recognition using smartphone data (HAPT-Recognition) to evaluate our algorithms. The experimental results on these datasets demonstrate that the proposed approach effectively removes private information from the datasets while allowing non-private information to be mined effectively.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Embargo reason
  • Ongoing Research
Embargo date range
  • 2021-08-29 to 2022-09-29

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details AlsulaimawiZahir2021.pdf 2021-08-28 Public Download