A Two-Tier Secure Federated Learning Framework with Lightweight Cryptography for Edge-Cloud Collaboration
Shahid Latif, Djamel Djenouri · 2025
This paper deals with scalability and security for large-scale deployments of Federated Learning (FL) and proposes a two-tier secure FL architecture that incorporates edge servers as intermediate aggregators between clients and the cloud. This hierarchical framework enhances scalability and minimizes cloud communication by utilizing edge-level aggregation. Security is maintained with lightweight cryptographic protocols, including X25519 for key exchange, HKDF with BLAKE2b for key derivation, and ChaCha20-Poly1305 for authenticated encryption, ensuring end-to-end confidentiality and integrity of model updates. An experimental evaluation on the MNIST dataset across three deployment scenarios (ranging from 9 to 100 clients) reveals improvements in model accuracy from 93.90% to 96.04%. Communication overhead to the cloud was reduced by up to 90%, and cryptographic overhead remained under 5.3ms per operation. Additionally, the architecture achieved 100% resistance to model poisoning, 97.8% prevention of gradient leakage, and 100% confidentiality preservation. The proposed approach outperformed existing state-of-the-art methods by a margin of 1.27% to 7.64% and demonstrated a strong balance between performance, security, and communication efficiency in edge-cloud FL environments.