2026-05-22T14:06:25Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2023972025-08-18T22:22:09Zoai_pmh:alloai_pmh:repo_items

202397 https://hdl.handle.net/2286/R.2.N.202397 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 155 pages Doctoral Dissertation Academic theses en Chen, Yitao Zhao, Ming Li, Baoxin Candan, Kasim Selcuk Wei, Hua Han, Kyungtae Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Engineering Artificial intelligence has enabled applications such as natural language processing and autonomous driving, leveraging deep neural networks (DNNs) in the cloud. However, growing data and computational demands expose limitations: 1) Scalability. The cloud struggles with exploding edge data; 2) Responsiveness. Latency and outages hinder real-time performance; 3) Privacy. Safeguarding sensitive data and models in the cloud is challenging. Advances in edge devices have made EdgeAI--machine learning on the edge, a promising solution. 1) Scalability. Edge devices provide compute for the data explosion. 2) Responsiveness. On-device processing reduces latency and avoids outages. 3) Privacy. Local data and models reduce cloud-related threats. However, EdgeAI also faces important challenges: First, limited and diverse edge resources; Second, varied application needs; Finally, edge-cloud collaboration raises privacy concerns. This dissertation optimizes EdgeAI for real-world resource and data constraints, achieving accuracy, real-time responsiveness, and privacy. It investigates: 1) What are the computational capabilities of edge devices for supporting machine learning? 2) How to design computational tasks for the edge to meet responsiveness requirements? and 3) How to improve privacy while allowing effective edge-cloud collaboration? First, this dissertation extends TensorFlow for training on Android and benchmarks DNNs on edge devices, showing they can support DNN inference and training with reasonable performance. Second, it proposes deployment-aware specialization and system-level co-design, and demonstrates its effectiveness through a case study on LiDAR-based crowd counting. The proposed method achieves 99.97% human classification accuracy, outperforms the state of the art (SOTA) by up to 51.37%, reduces crowd counting errors by up to 90.44%, and meets real-time responsiveness at 17.42 ms per sample. Third, it proposes confidence-based federated distillation to enhance privacy, and a data fusing and cross-modal framework for addressing training-time modality incompleteness. The proposed federated distillation method shares only model outputs between edge and cloud and improves accuracy over SOTA by up to 11.3%. The proposed data fusing and cross-modal framework constructs a pseudo-multimodal dataset and trains a cross-attention multimodal transformer, achieving 96.7% accuracy and outperforming SOTA by up to 6.18%. Computer Engineering Optimizing Edge Machine Learning for Real-world Resource and Data Constraints