Towards Trustworthy Learning in Temporal Learning Environments

This thesis explores how to build trustworthy machine learning systems that learn and adapt over time. As machine learning moves beyond static benchmarks and into real-world settings, where data distributions shift, environments change, and objectives evolve, it is more important and difficult to ensure consistent and reliable behavior. Yet, most existing methods for safety and robustness are designed for models trained on fixed datasets that do not account for the dynamics of temporal learning, where models must continually update, interact with changing environments, or operate under evolving constraints. To address this gap, the thesis investigates three representative settings of temporal learning: continual learning, reinforcement learning, and constraint-driven optimization. It begins by showing how continual learning systems that rely on generative replay can be subtly and persistently compromised. A new data poisoning technique is introduced that embeds backdoors into training data, taking advantage of the limited capacity of generative models. The attack causes models to forget previously learned tasks over time, while maintaining strong performance on current tasks. This underscores a critical vulnerability of continual learners: they can be silently compromised, with the effects only emerging after task transitions—when it is too late to recover. The thesis then turns to reinforcement learning, focusing on a practical variant known as offline-to-online RL. In this setting, agents are first trained on pre-collected data and later fine-tuned online in real environments. A novel attack is presented that perturbs reward signals in the offline data just enough to remain undetected during offline training, but causes significant performance degradation once the agent is deployed online. This exposes a blind spot in current evaluation practices, which often equate strong offline performance with real-world reliability. Finally, the thesis explores how adaptive data selection can enhance learning in systems governed by universal constraints. It introduces a reinforcement learning-based framework that dynamically selects training inputs in response to the evolving state of the model. This approach is applied to Lyapunov neural networks—aiming to certify the stability of dynamical systems, and physics-informed neural networks—enforcing physical consistency through partial differential equations. By tailoring the input distribution to the model’s learning progress, the method improves both training efficiency and constraint satisfaction compared to traditional static or heuristic sampling strategies. Taken together, these contributions reveal how temporal learning systems can fail in subtle, time-dependent ways, and how their performance and reliability can be improved through adaptive, temporally-aware methods. Rather than relying solely on static safeguards, the thesis emphasizes the need to understand and anticipate how learning unfolds over time. It offers both a critique of the limitations in current trustworthiness frameworks and a path forward for designing more resilient and responsive models capable of operating effectively in dynamic, real-world environments.

Author dc:creator

• Siteng Kang (23291689)

• Computer Science

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451434.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451434