FLSecure: A hybrid framework with blockchain and multi-TEE parallel execution for secure federated learnings

Abstract

Federated Learning (FL) enables decentralized model training on private data but faces challenges such as single points of failure, adversarial attacks, and limited scalability. While blockchain-based FL frameworks address some of these issues by decentralizing model aggregation, they often suffer from computational inefficiencies and insufficient confidentiality during model updates. This paper presents FLSecure, a hybrid framework that integrates blockchain and Trusted Execution Environments (TEEs) to improve the privacy, security, and scalability of FL systems. By utilizing TEEs, FLSecure ensures secure aggregation of local model updates within isolated hardware environments, preventing unauthorized access and ensuring data integrity. Blockchain provides decentralized consensus and a tamper-proof audit trail, removing the dependency on centralized servers and enhancing transparency. FLSecure introduces a multi-TEE parallel execution strategy, which partitions global aggregation tasks into subtasks and distributes them across multiple TEEs for concurrent execution. This strategy mitigates TEE memory constraints and blockchain transaction bottlenecks, thereby enhancing scalability, reducing computational overhead, and supporting large-scale deployments. Experimental results demonstrate that FLSecure outperforms existing privacy-preserving FL frameworks in computational efficiency and overhead reduction. Furthermore, it exhibits robust resilience against adversarial threats, including backdoor and Byzantine attacks, offering a comprehensive solution to the security and scalability challenges in federated learning.