Resource-Efficient Collaborative Training and Inference of Foundation Models in Edge-AI

The convergence of Edge Artificial Intelligence (Edge-AI) and foundation models marks a transformative paradigm shift in the design of intelligent systems. Edge-AI enables computation to be performed closer to data sources and across distributed network edges, offering significant benefits in latency reduction, privacy preservation, and real-time decision-making. Meanwhile, foundation models, pre-trained on large-scale and diverse datasets, exhibit unprecedented generalization capabilities across a wide range of downstream tasks, including natural language processing, computer vision, and multimodal applications. The integration of foundation models into Edge-AI environments promises to unlock tremendous potential for intelligent applications such as autonomous driving, smart cities, and Industry 4.0. However, the substantial computation, memory, and communication demands of deploying foundation models in Edge-AI networks present critical challenges for resource-constrained edge environments. These challenges necessitate the development of resource-efficient, system-level methodologies for collaborative model training and inference. This thesis addresses the aforementioned challenges by investigating how heterogeneous edge devices and edge servers can jointly train and leverage foundation models under resource constraints. The research identifies three core research problems: 1) how to collaboratively train large-scale foundation models without overwhelming edge device resources, 2) how to effectively coordinate self-interested edge devices in the process of federated foundation model fine-tuning, and 3) how to deliver personalized, low-latency synergistic edge inference services from the perspective of the Edge-AI market. To address the first research problem, this thesis proposes DeepFusion, a scalable federated knowledge distillation framework that enables heterogeneous edge devices to transfer lightweight local model knowledge into a global Mixture of Experts-based foundation model. This approach removes the burden of hosting entire foundation models on resource-limited devices, while still aggregating diverse local expertise into a unified global foundation model. To address the second research problem, this thesis introduces PRINCE, a novel incentive mechanism tailored for multi-tenant Split Federated Learning (SFL), enabling resource-efficient fine-tuning of foundation models across multiple downstream tasks. By modeling self-interested device participation as a multi-leader multi-follower Stackelberg game, PRINCE allows SFL tenants to strategically allocate incentives, thus attracting high-quality edge contributions while ensuring bias-resilient aggregation and provable convergence of the global foundation model. To address the third research problem, this thesis develops AERIA, a market-oriented on-demand synergistic edge inference framework for foundation models. Leveraging auction-driven resource allocation, AERIA regulates market interactions among AI service providers, users, and edge infrastructure operators, thereby ensuring fairness, incentive compatibility, and revenue competitiveness in auction outcomes, while simultaneously accommodating the personalized inference demands of diverse AI users. Collectively, these methodologies address distinct stages of the Edge-AI lifecycle and form an end-to-end pipeline encompassing foundation model training, fine-tuning, and inference service provisioning. The proposed approaches are validated through extensive real-world simulation and trace-driven experiments, demonstrating significant improvements in system scalability and resource efficiency compared to state-of-the-art solution approaches. Overall, this thesis contributes novel system designs for resource-efficient collaborative training and inference of foundation models in Edge-AI, advancing the development of scalable, adaptive, and trustworthy intelligent services at the network edge.<p></p>

Author dc:creator

• Songyuan Li (21053258)

• Edge-AI
• Foundation Models
• Federated Knowledge Distillation
• Split Federated Learning
• Synergistic Edge Inference
• Incentive Mechanisms

Statement dc:rights

• CC BY
• Open Access after 2027-07-26

Identifier

10779/exe.31096063.v1

OAI identifier oai:identifier

oai:figshare.com:article/31096063