Domain Adaptation for Fair and Robust Machine Learning
Abstract
Deep learning society has grown tremendously in recent years, leading to increasing demand for large amounts of data. However, it is difficult to gather enough labeled data similar to unlabeled real-world test data. We have to apply the domain adaptation to overcome this limitation. Domain adaptation learns by taking into account differences in probability distributions between different domains, so it outcomes a model that can be applied to both domains. In this dissertation, we first focus on developing domain adaptation methods in terms of learning appropriate representation. We propose a class-conditional domain invariant learning method that can learn a feature space in which features in the same class are expected to be mapped nearby. Improving the performance of domain adaptation has an expected impact that the application area of machine learning will be widened because a large number of unsupervised data can be used for machine learning. In addition to these direct expected impacts, domain adaptation techniques can be used in many aspects of deep learning. The key characteristic of domain adaptation is that it builds a machine learning model that can be applied to two different distributions. Therefore, domain adaptation based approaches can be widely used in machine learning areas that handle two different distributions, such as defense against adversarial attacks or fair machine learning. Robustness of deep learning is important since it can induce various security and safety problems. We focus on modeling deep networks that are robust against adversarial attacks by using domain adaptation methods. Considering normal samples as the source domain and adversarial samples as target domain, we suggest novel Wasserstein distance-based domain adaptation method that can enhance model robustness. As the increase in machine learning influences decisions about important aspects of our lives, the demand for a fair machine learning model has increased rapidly in recent years. We investigate a methodology for developing a fair classification model for data with limited or no labels, by transferring knowledge from another data domain where information is fully available, which is done by controlling the Wasserstein distances between relevant distributions. In real-world applications of machine learning, there are many cases where we need to handle data from various distributions. In these cases, domain adaptation can be the right solution because it considers the difference of probabilistic distributions of two domains. We expect, our algorithms can be helpful to overcome these limitations and be applied to build fair and robust models.
Keywords
Not available
Specialisations
Similar theses
- A Study on Robustness, Adaptation, and Generalization for Visual and Multimodal Learning Xiasi Wang · 2026
- A Comprehensive Framework for Targeted Transfer Learning: Privacy-Preserving, Secure, and Efficient Domain Adaptation Sayyed Farid Ahamed · Old Dominion University · 2026
- Towards transfer learning in e-discovery :finding the optimal classifier and evaluating domain adaptation methods J.L. Pebesma · 2020
- Deep Reinforcement Learning for Intrusion Detection and Response in Software-Defined Networks Trung Phan Van · Chemnitz University of Technology · 2026
- Evaluating Self-Supervised Learning for Network Intrusion Detection Bryan A. Marin Samano · 2026
- Multi-Representation Learning for Domain Adaptation in Machine Learning Applications Victor Akihito Kamada Tomita · Universidade de São Paulo · 2025
