Scholarlink

<p>Federated Learning (FL) is gaining traction as a method for training machine learning models in distributed environments while preserving data privacy. However, applying FL in cross-silo organizations, such as healthcare or finance, introduces challenges related to privacy and verifiability due to the sensitive nature of data and strict regulatory requirements. This paper presents V-FLEX, a novel framework for verifiable cross-silo federated learning that decentralizes trust by reducing reliance on a single entity. V-FLEX ensures the secure aggregation of encrypted updates and verifies the integrity of both local and global model updates. By leveraging Trusted Execution Environments (TEEs) for proof mechanisms and threshold cryptography, V-FLEX provides strong defenses against inference attacks, collusion, and data tampering. We conduct a comprehensive security analysis under a semi-honest threat model to validate its security guarantees. Our primary objective is to achieve verifiable aggregation of federated model updates while maintaining competitive accuracy and manageable computational overhead. Experimental results show that V-FLEX maintains model accuracy comparable to traditional FL methods, achieving a final accuracy within 0.5% of FedAvg for both IID and non-IID datasets. Despite integrating cryptographic mechanisms such as homomorphic encryption and secure verification, V-FLEX incurs only moderate computational and communication overhead, demonstrating its practical feasibility.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

V-FLEX: Verifiable cross-silo federated learning using trusted execution environment

  • 期刊论文

  • 发表日期：

    2025年7月26日

查看全文
  • Jaouhara Bouamama,
  • Yahya Benkaouz,
  • Mohammed Ouzzif

摘要

Federated Learning (FL) is gaining traction as a method for training machine learning models in distributed environments while preserving data privacy. However, applying FL in cross-silo organizations, such as healthcare or finance, introduces challenges related to privacy and verifiability due to the sensitive nature of data and strict regulatory requirements. This paper presents V-FLEX, a novel framework for verifiable cross-silo federated learning that decentralizes trust by reducing reliance on a single entity. V-FLEX ensures the secure aggregation of encrypted updates and verifies the integrity of both local and global model updates. By leveraging Trusted Execution Environments (TEEs) for proof mechanisms and threshold cryptography, V-FLEX provides strong defenses against inference attacks, collusion, and data tampering. We conduct a comprehensive security analysis under a semi-honest threat model to validate its security guarantees. Our primary objective is to achieve verifiable aggregation of federated model updates while maintaining competitive accuracy and manageable computational overhead. Experimental results show that V-FLEX maintains model accuracy comparable to traditional FL methods, achieving a final accuracy within 0.5% of FedAvg for both IID and non-IID datasets. Despite integrating cryptographic mechanisms such as homomorphic encryption and secure verification, V-FLEX incurs only moderate computational and communication overhead, demonstrating its practical feasibility.