Advancing Edge AI: Frameworks and Strategies for Trustworthy and Efficient Collaborative Learning

Artificial Intelligence is increasingly shifting from centralized cloud infrastructures to distributed edge environments, yet current models still treat edge devices mainly as passive data sources. This creates bandwidth bottlenecks, privacy risks, and limited support for collaborative training across heterogeneous devices. This thesis extends the cloud-edge continuum to enable seamless AI execution across cloud, edge, and client resources, using federated learning as both an application domain and an analytical lens. It addresses four core challenges: orchestration, accessibility, trustworthiness, and production readiness. The work introduces the Cloud-Edge-Client Continuum architecture for temporary clusters in safety-critical settings; FLAT, a zero-install browser-based federated learning approach built on ONNX and WebAssembly; and extensions to Flower for continuum-aware operation. It further improves trustworthiness through homomorphic encryption, FedObj for detecting dishonest clients, and HEXA for efficient federated fine-tuning under high variability. Finally, it focuses on the gap between research and production, defines evaluation metrics, and presents FLeeT, a Kubernetes-native federated learning platform benchmarked against Flower and NVIDIA FLARE, exposing key operational gaps.