The rapid advancement of machine learning, particularly over-parameterized deep neural networks (DNNs), has led to significant progress across diverse domains. While the over- parameterization of DNNs gives them the power to capture complex mappings between input data points and target labels, in real-world challenges, they can inevitably be exposed to unseen out-of-distribution (OoD) examples that deviate from the training distribution. This raises critical concerns around robustness, adaptiveness, and trustworthiness of such models when transferring knowledge from the training domains to unseen test domains.In this thesis, we propose three different methods targeting the robustness and adaptiveness of machine learning models. First, to address agnostic data corruption in the source domain, we propose a simple and computationally efficient unsupervised domain adaptation (UDA) approach that enables parallel training of ensemble models. The learning framework we proposed can be flexibly combined with available UDA approaches that are orthogonal to our work to improve their robustness under corrupted data. Second, with the rise of large language models (LLMs) pre-trained on vast, web-sourced datasets spanning multiple domains, which led to a surge of interest in adapting these models to a wide range of downstream tasks. However, the real-world corpora used in the pre-training stage often exhibit a long-tail distribution, where knowledge from less frequent domains is underrepresented. As a result, LLMs failed to give correct answers for queries sampling from the long-tail distributions. To solve this problem, we propose a reinforcement learning-based dynamic uncertainty ranking method for retrieval-augmented ICL with a budget controller. The system adjusts the ranking of retrieved samples based on LLM feedback, promoting informative and stable examples while demoting misleading ones. Third, while the neighborhood community DA aims to ensure model robustness by maintaining high performance on OoD samples from target domains with domain shifts, out-of-distribution (OoD) detection focuses on model reliability by identifying samples that exhibit semantic shifts. To bridge a critical research gap of OoD detection and federated learning (FL), we propose a privacy-preserving federated OoD synthesizer that exploits data heterogeneity to enhance out-of-distribution (OoD) detection across clients. This approach enables each client to benefit from external class knowledge shared among non-IID participants, without compromising data privacy.The model adaptation process can also introduce a new challenge, which is the risk of unauthorized reproduction or intellectual property (IP) theft, especially for high-value models. To enhance the trustworthiness of models, we introduce two methods for model watermarking. The first is an OoD-based watermarking technique that eliminates the need for training data access, making it suitable for scenarios with strict data confidentiality. The method is both sample-efficient and time-efficient while preserving model utility. The second technique targets federated learning, enabling both ownership verification and leakage tracing, transitioning FL model use from anonymity to accountability.
Read
