Securing Federated LLM Training through Blockchain-enabled Network Architecture

Federated Large Language Model (LLM) training enables multiple participants to collaboratively develop AI models without sharing raw data, thereby maintaining privacy. However, this decentralized method introduces security challenges, including potential data integrity issues, model poisoning threats, and trust concerns among participants. To address these issues, this paper proposes a blockchain-enabled network architecture designed to enhance the security and reliability of federated LLM training.By integrating blockchain technology, the proposed framework provides transparency, immutability, and decentralized trust mechanisms. The blockchain maintains comprehensive records of all model updates, enabling verifiable, tamper-resistant auditing, and supporting consensus-based validation to identify and prevent malicious contributions. This architecture aims to strengthen defenses against adversarial attacks while promoting accountability among participants.The study assesses the system's performance, scalability, and security measures, demonstrating its effectiveness in mitigating common risks associated with federated learning. A detailed diagram illustrates the architecture, emphasizing key components such as distributed nodes, consensus protocols, and secure update mechanisms. Additionally, references to relevant literature underpin the theoretical foundation, covering advancements in federated learning, blockchain security, and adversarial robustness.This innovative approach aims to bridge privacy-preserving AI training with robust cybersecurity principles, offering a scalable solution for secure, collaborative development of large language models. Future research may focus on optimizing this architecture for practical deployment across various industries.

کلیدواژه ها:

Federated Learning ، Large Language Models (LLMs) ، Blockchain ، Model Poisoning ، Decentralized Trust ، Byzantine Fault Tolerance ، IPFS ، Smart Contracts