Achieving Differential Privacy and Fairness in Machine Learning

<p>Machine learning algorithms are used to make decisions in various applications, such as recruiting, lending and policing. These algorithms rely on large amounts of sensitive individual information to work properly. Hence, there are sociological concerns about machine learning algorithms on matters like privacy and fairness. Currently, many studies only focus on protecting individual privacy or ensuring fairness of algorithms separately without taking consideration of their connection. However, there are new challenges arising in privacy preserving and fairness-aware machine learning. On one hand, there is fairness within the private model, i.e., how to meet both privacy and fairness requirements simultaneously in machine learning algorithms. On the other hand, there is fairness between the private model and the non-private model, i.e., how to ensure the utility loss due to differential privacy is the same towards each group.</p> <p>The goal of this dissertation is to address challenging issues in privacy preserving and fairness-aware machine learning: achieving differential privacy with satisfactory utility and efficiency in complex and emerging tasks, using generative models to generate fair data and to assist fair classification, achieving both differential privacy and fairness simultaneously within the same model, and achieving equal utility loss w.r.t. each group between the private model and the non-private model. </p> <p>In this dissertation, we develop the following algorithms to address the above challenges. </p> <p>(1) We develop PrivPC and DPNE algorithms to achieve differential privacy in complex and emerging tasks of causal graph discovery and network embedding, respectively.</p> <p>(2) We develop the fair generative adversarial neural networks framework and three algorithms (FairGAN, FairGAN+ and CFGAN) to achieve fair data generation and classification through generative models based on different association-based and causation-based fairness notions.</p> <p>(3) We develop PFLR and PFLR* algorithms to simultaneously achieve both differential privacy and fairness in logistic regression.</p> <p>(4) We develop a DPSGD-F algorithm to remove the disparate impact of differential privacy on model accuracy w.r.t. each group. </p>

Name thesis:degree_name

Doctor of Philosophy in Engineering (PhD)

Level thesis:degree_level

Dissertation

Year dc:date.available

2021

Author dc:creator

• Xu, Depeng

Advisor dc:contributor.advisor

• Wu, Xintao

Contributors dc:contributor

• Panda, Brajendra N.
• Li, Qinghua

• algorithmic fairness
• differential privacy
• machine learning
• Databases and Information Systems
• Graphics and Human Computer Interfaces
• Information Security
• Numerical Analysis and Scientific Computing
• Systems and Communications

Repository record dc:identifier

https://scholarworks.uark.edu/etd/3960

OAI identifier oai:identifier

oai:scholarworks.uark.edu:etd-5510