Content area
In 2016, Federated Learning (FL) was introduced as a new privacy-preserving distributed machine learning paradigm, functional especially in edge computing environments. However, this paradigm does not take into consideration the growing necessity to guarantee that Human-in-theLoop (HITL) methodologies are robustly integrated into AI systems, as standardized integration approaches remain absent. Therefore, to bridge this gap, this thesis proposes a comprehensive, modular reference architecture for introducing HITL strategies into FL workflows at the edge, using the NOUS project (an EU initiative for next-generation cloud services) as a use case.
Our work begins with a systematic literature review and gap analysis to characterize the current practices and identify the core challenges in this research domain. From this foundation, we derive architectural high-level goals alongside functional and non-functional requirements. From this, we then present the HITL-FL framework, which utilizes C4 models (System Context, Container, Component). This framework includes a dedicated Human Oversight & Interaction Hub, secure annotation interfaces, task routing, ethical validators and operational loops for active learning and model governance.
We validate the architecture through three pillars: systematic compliance analysis against our requisites; mapping using the NOUS project and its Use Case #2 (Energy Prediction & Data Lifecycle Management); and qualitative expert reviews with NOUS architects. Validation results showcase strong support for comprehensive human oversight and alignment with trustworthiness, privacy, security, and usability requirements while also uncovering limitations and design challenges for future work.
Through this process, this research delivers a foundational blueprint for building humancentric, transparent, federated AI systems that foster a responsible collaboration between humans and AI in complex edge ecosystems.