CAFEIN is a federated learning platform connecting research institutions and industry partners to train AI models on sensitive data without moving it off-site. Built on CERN’s on-premises cloud with Kubernetes, it enables scalable, privacy-preserving training across diverse domains.
This talk outlines CAFEIN’s secure, scalable architecture on Kubernetes, including design decisions enabling federated learning at scale under strict privacy constraints. It also covers integration of tools like Argo CD, Falco, and Prometheus/Grafana to orchestrate distributed training across institutions. It explains how deployment automation and policy controls (Vault for secrets, Kyverno for policies) protect data and manage resources. The session addresses multi-site AI workflow challenges such as network isolation, single sign-on (Keycloak), and monitoring. Attendees will learn practical techniques for orchestrating secure, multi-institution federated learning workflows.