On the operational challenges of federated learning: sustainability, cost efficiency, and regulatory compliance in multicloud infrastructures

Federated Learning has emerged as a transformative paradigm for privacy-preserving collaborative model training, enabling geographically distributed institutions to build shared intelligence without ever exposing their local data. Its appeal is particularly strong in regulated domains such as healthcare and finance, where data sovereignty and legal accountability are non-negotiable. Yet as FL matures from research prototype to production infrastructure, a deeper challenge surfaces: real-world deployments increasingly span multiple cloud providers simultaneously, introducing a layer of operational complexity that existing frameworks are not equipped to handle. Managing heterogeneous billing models, fragmented resource telemetry, and provider-specific APIs across AWS, Azure, and GCP creates inefficiencies that translate directly into financial waste, unnecessary carbon emissions, and reduced organizational control, undermining the very sustainability goals that motivate cloud-native AI adoption. Beyond cost and sustainability, regulated FL deployments face an equally critical gap at the compliance layer. Collaborating across institutional and national boundaries requires demonstrable adherence to data protection regulations, contractual obligations, and runtime integrity standards, requirements that current orchestration solutions largely ignore. The result is a fragmented landscape where privacy-preserving algorithms coexist with legally unverifiable infrastructure, audit trails are incomplete or absent, and the environmental cost of distributed training remains invisible and unmanaged. Addressing these gaps in isolation is insufficient; what is needed is a unified approach that treats financial sustainability, environmental responsibility, and regulatory compliance as inseparable, co-equal objectives of the orchestration layer itself. This thesis presents two novel, complementary contributions that jointly address these open problems. The first, ECOFED, tackles the multicloud sustainability challenge by introducing a FinOps-driven orchestration middleware that brings financial observability and dynamic resource governance to federated training workloads. By treating environmental sustainability as a direct consequence of sound financial management, minimizing idle cloud resources reduces both cost and carbon footprint simultaneously, allowing the framework to enable organizations to run FL pipelines across heterogeneous providers in a way that is both economically viable and ecologically responsible. The second contribution, ECOFED-X, builds on this foundation by adding a compliance-aware, zero-trust orchestration layer designed for safety-critical applications. It ensures that participation in each training round is governed by verifiable legal, contractual, and runtime conditions, and that every decision is permanently recorded for regulatory auditability without ever centralizing sensitive data. Taken together, these contributions represent a novel and holistic step forward for Federated Learning infrastructure: from a model-centric paradigm operating in isolated cloud silos, toward a governance-aware, sustainability-driven, and legally accountable system capable of meeting the demands of modern regulated environments. Experimental validation in realistic multicloud scenarios confirms that strong operational and compliance guarantees can be achieved with negligible impact on model quality, positioning this work as a concrete path toward responsible, scalable, and green federated learning.