FLChain-LLM: A Blockchain-Enabled Framework for Training Large Language Models

Document Type

Conference Proceeding

Source of Publication

2025 7th International Conference on Blockchain Computing and Applications (BCCA)

Publication Date

10-14-2025

Abstract

The training of large language models (LLMs) on decentralized and privacy-sensitive data poses significant challenges related to trust, coordination, and security. To address these issues, we propose FLChain-LLM, a blockchain-enabled federated learning (FL) framework designed for collaborative fine-tuning of pre-trained LLMs across distributed clients. Our framework integrates smart contracts to automate client participation, secure model update verification, and reward contribution, while ensuring transparency and immutability through blockchain records. We employ GPT-2 as a base model and fine-tune it using the StackOverflow dataset partitioned by user, simulating a realistic FL environment. Experimental results show a reduction in global model perplexity from 34.7 to 27.8 over 10 rounds. The token-based incentive mechanism fairly rewards clients based on local loss and participation quality. FLChain-LLM demonstrates strong resilience to nonIIID data and malicious clients while preserving model utility. Our work highlights the potential of blockchain as a trust infrastructure for scalable, privacy-preserving LLM adaptation in decentralized settings, such as edge computing or crossorganization collaborations.

Publisher

IEEE

First Page

378

Last Page

384

Disciplines

Computer Sciences

Keywords

Privacy-Preserving Technologies in Data, Blockchain Technology Applications and Security, Blockchain Technology in Education and Learning

Indexed in Scopus

no

Open Access

no

Link to Full Text

Share

COinS