Date of Award
Spring 5-6-2025
Degree Type
Dissertation/Thesis
Degree Name
Master of Science in Information Technology
Department
Information Technology
Committee Chair/First Advisor
Liang Zhao
Second Advisor
Shaoen Wu
Third Advisor
Bobin Deng
Abstract
Federated learning (FL) is a novel paradigm that enables the training of a global machine learning (ML) model across distributed devices by exchanging model parameters instead of raw data in the training process. Internet of Things (IoT) devices typically operate with limited resources, have weaker security protections, and are more vulnerable to potential thermal stress (TS). Current evaluations of FL are mostly conducted through simulations of multiple clients on a single device. However, there remains a gap in understanding how FL performs under TS in real-world, low-power IoT environments. Conformal prediction (CP) is an effective method for quantifying uncertainty in ML models. However, the fundamental assumption of CP, that data are exchangeable, does not inherently hold in FL. In this paper, we present a unified evaluation of FL considering these challenges. First, we evaluated the performance of FL under TS on a homogeneous real-world distributed IoT system. Second, we extended this evaluation to a larger-scale heterogeneous system. Third, we examined the performance of CP under FL using a single-node simulated distributed system. To the best of our knowledge, this work is among the first to explore these issues. For the TS evaluation, we conducted comprehensive experiments using the CIFAR-10 and MNIST datasets and measured performance metrics like training time, CPU and GPU utilization rates, temperature, and power consumption. We varied the proportion of clients exposed to TS across experimental groups and systematically quantified the effects of TS on low-end IoT-based federated learning systems (FLS). For CP evaluation, we assessed its performance on an FLS performing a regression task on the California housing dataset. The results show that TS has a significant impact on IoT-based FLS, as both global model performance and device operation degrade even when only a small fraction of clients are affected. We also found that the effectiveness of CP deteriorates when applied in FLS.