TrusTEE: Storage-Centric Secure Federated Learning with Trusted Execution and Policy Enforcement

Abstract

: Federated Learning enables collaborative model training without sharing raw user data, yet remains vulnerable to privacy threats via shared gradients, model updates, and the hardware in which aggregation runs. Users must trust that centralised servers correctly aggregate updates and enforce data access policies. This paper introduces TrusTEE, an end-to-end hardware-backed confidential federated learning system that combines cloud Trusted Execution Environments (TEEs) with user-controlled, policy-enforced object storage. Unlike prior TEE-based federated learning work, TrusTEE treats cloud object storage as the primary communication substrate and trust boundary, using a minimal TEE-based aggregator and cloud-native role-based access control (RBAC) to separate roles. The system ensures that (1) raw client updates are never visible to the coordinator, (2) aggregation occurs entirely in hardware-encrypted memory, and (3) access to updates, aggregates, and global weights is enforced via strict identity-bound storage policies. We evaluate TrusTEE using the FLAIR dataset and show that confidential aggregation can be achieved while preserving standard federated learning dynamics.