Large language models (LLMs) have achieved breakthrough advancements in natural language processing, yet privacy preservation during fine-tuning remains a critical challenge for real-world deployment. Current federated LLM fine-tuning frameworks avoid raw data sharing through collaborative training but still face risks of gradient leakage attacks, which can expose sensitive client information. Meanwhile, model divergence from data heterogeneity and performance drops from traditional differential privacy in resource-constrained scenarios remain issues. To address these challenges, this paper proposes DecFLLM, a privacy-preserving framework for federated LLM fine-tuning. DecFLLM innovatively decomposes adapter parameters into two components: global universal adapters (capturing cross-client shared knowledge) and local private adapters (capturing client-specific private knowledge). This thwarts gradient leakage attacks by preventing exposure of sensitive gradient info, while retaining model personalization and robustness amid high data heterogeneity. This work establishes a novel paradigm for privacy-performance co-optimization in FL-based LLM fine-tuning, advancing the development of privacy-aware machine learning systems.