Divide And Conquer: Efficiently Decoupling Consensus And Divergence For Federated Large Language Model Fine-Tuning
Keywords: Federated Learning, Large Language Model
Abstract: Federated Learning provides an efficient framework for fine-tuning Large Language Models (LLMs) on diverse private datasets, addressing the growing scarcity of publicly available training data while maintaining data privacy. However, in practice, client data typically spans multiple domains, posing significant challenges for the global model’s generalization capabilities. To address this issue, we introduce a novel framework, **Fed**erated **C**onsensus-**D**ivergence **D**ecoupling for LLM Fine-Tuning (**FedCDD**), designed to enhance global model performance in such heterogeneous environments. Our framework introduces a mechanism for consensus aggregation and divergence alignment, decoupling client updates into “consensus” and “divergence” parts. This allows the LLM to maintain a unified consensus while accommodating domain-specific divergences. Additionally, we employ a Gaussian-Noise Mask to regulate local model uploads, preventing the LLM from overfitting to domain-specific knowledge. Experimental results on heterogeneous datasets demonstrate the superiority of our approach over existing methods. The code is anonymously available at https://anonymous.4open.science/r/FedCDD-5DA6.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10214
Loading