LLM‐Powered Data Synthesis on Blockchain for Fusion Isomerism Learning in Heterogeneous Federated Systems

Abstract

Severe data starvation, architectural heterogeneity and Byzantine vulnerabilities fundamentally impede the deployment of robust multiclass classification models in decentralised edge environments. To address these intertwined challenges, we propose fusion isomerism learning (FusionIL) , a secure and domain‐agnostic framework that synergistically orchestrates isomerism large language models (IsoLLMs), consortium blockchain technology and heterogeneous federated learning. Unlike conventional approaches, FusionIL leverages a trusted committee of permissioned nodes to generate high‐fidelity synthetic data via feature‐driven textual prompts, forming IsoLLMs that capture structural heterogeneity across models. To prevent malicious data pollution, these generated samples undergo rigorous quantitative validation and are securely anchored via the practical Byzantine fault tolerance (PBFT) consensus mechanism. Furthermore, to mitigate the risk of generative model collapse, permissionless edge clients construct “synthetic interleaved datasets” by merging these verified synthetic samples with their private data under strict minority proportions. Subsequently, the framework employs the decentralised heterogeneous integration algorithm (DHIA) to mathematically reconcile and aggregate gradient updates from structurally diverse local models. This architectural innovation effectively eliminates the rigid homogeneity constraints of traditional federated networks. Extensive empirical evaluations across multiple benchmark datasets demonstrate that the synthetic interleaving strategy significantly enhances data diversity and global convergence. Consequently, FusionIL achieves competitive or leading performance against state‐of‐the‐art centralized and federated baselines in classification accuracy and system resilience, establishing a robust paradigm for secure, heterogeneous collaborative learning.