Trustworthy AI Under Distribution Shifts: Enhancing Generalization and Fairness in Deep Learning

Recent advances in machine learning and deep learning have achieved remarkable success across a wide range of applications, yet they largely rely on data-driven paradigms that assume the training and test data share similar distributions. In real-world scenarios, however, this assumption often fails. Distribution shifts between training and deployment environments or among population subgroups can significantly compromise the performance and reliability of AI systems. Moreover, privacy constraints frequently prevent direct access to source data, and ethical concerns surrounding fairness raise additional challenges for the trustworthy use of AI in sensitive domains such as healthcare and finance. These factors together highlight the urgent need for a deeper and more principled understanding of trustworthy AI under distribution shifts. This thesis examines the influence of distribution shifts along two complementary dimensions of trustworthy AI: generalizability and fairness. On one hand, we focus on the noisy and unreliable signals caused by distribution shift, which place learning in a weak-supervision regime and thereby limit AI systems’ generalizability. On the other hand, we focus on the distribution shifts that arise from structural data imbalance or scarcity rooted in historical and societal factors, which give rise to socially unfavorable model behaviors and, in turn, exacerbate fairness issues in the decision-making process. From the generalizability perspective, the thesis studies the domain adaptation problem under data privacy constraints, formulated as the Source-Free Domain Adaptation (SFDA) problem. SFDA is examined in both classification and regression tasks, with a focus on pseudo-label noise mitigation and uncertainty modeling, respectively. For the classification problem, a noise-robust loss–based noise-and-variance control method (NVC-LLN) is proposed to mitigate the unbounded label noise inherent in SFDA. For the regression task, a histogram-based uncertainty reconstruction framework, MERCI, is proposed to refine continuous supervision signals and align feature presentations under distribution shift. For the fairness problem, this thesis focuses on the multiple-subgroup situation, ranging from a single sensitive attribute with multiple subgroups to multiple sensitive attributes with intersectional bias, covering both unfairness discovery and mitigation. For subgroup discovery, a Bias-Guided Generative Network (BGGN) is proposed to identify underrepresented subgroups instead of relying on traditional search methods. For mitigation, a generalization error bound is derived under the PAC-Bayes framework, motivating a probabilistic-predictor-based bilevel optimization algorithm, FAMS, which achieves both fairness and accuracy across multiple subgroups under limited-data conditions. Collectively, these studies advance the development of responsible, adaptive, and human-centric AI systems. These systems are capable of performing reliably across diverse and dynamic environments.