Privacy-Preserving and Fairness in Machine Learning
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Machine learning is widely deployed in society, unleashing its magic in a wide range of applications following the progress of big data and computing power. However, society is beginning to realize that machine learning models designed to help human beings in various tasks may also have a negative impact on human beings, especially in terms of privacy and fairness. In terms of privacy, data are increasingly collected from human beings, and when these data are used for machine learning, data privacy might be compromised. In terms of fairness, machine learning, as a useful decision-making tool, is widely used to allocate resources and opportunities for humans. Many studies have shown that decisions made by these models may be biased against certain populations. Machine learning has passed the stage of only considering model performance, and ethical issues have a decisive impact on the use of machine learning. This thesis mainly studies how to design a fair and private machine learning model to foster private and fair machine learning and develops methods broadly covering different aspects of privacy and fairness to enhance the trade-off between fairness, privacy and model accuracy. Specifically, it makes the following contributions.
• We propose a correlation reduction scheme with feature selection - selecting features considering data correlation and utility. The proposed scheme involves five steps to manage the extent of data correlation, preserve privacy, and support accuracy in the model outputs.
• We present a framework of fair semi-supervised learning in the pre-processing phase, including pseudo labeling, re-sampling, and ensemble learning to improve accuracy and decrease discrimination. We also propose a framework of fair semi-supervised learning in the in-processing phase. The objective function includes a loss for both the classifier and label propagation and fairness constraints over labeled and unlabeled data.
• We study the balance between accuracy, privacy and fairness in deep learning by designing two different early stopping criteria to help analysts choose when to stop training a model to achieve their ideal trade-off.
• We investigate how adversarial examples will skew model fairness. We formulate the problem as an optimization problem: maximizing the model bias with the constraint of the number of adversarial examples and the perturbation scale.
Please use this identifier to cite or link to this item: