PureLLM: A Blockchain-Driven Decentralized PFL for Robust and Resource-Efficient Next-Gen LLMs

초록

Large language models (LLMs) trained on extensive publicly available datasets have achieved groundbreaking performance across diverse domains. However, both contemporary and future LLMs face four major challenges: (i) Data Availability, (ii) Data Bias and Quality, (iii) Computational Resources, and (iv) Data Privacy and Security. To address these challenges, we propose PureLLM, a blockchain-driven decentralized personalized federated learning (PFL) framework for the development of high-performance, resource-efficient next-generation LLMs. By leveraging decentralized federated learning, PureLLM enables collaborative training across diverse data sources without exposing sensitive user information. A blockchain-powered consensus mechanism ensures data integrity, security, and fairness, mitigating biases and enhancing trust in model updates. Additionally, our approach optimizes resource utilization by dynamically allocating computational workloads across distributed nodes, reducing energy consumption and improving scalability. Experimental evaluations demonstrate that PureLLM achieves superior model performance while maintaining strong privacy guarantees and reducing computational overhead. This framework paves the way for democratized and secure LLM training, fostering an inclusive AI ecosystem where individuals and organizations can contribute to and benefit from cutting-edge language models without compromising privacy or efficiency.