Instruction tuning has proven essential for enhancing the performance of large language models (LLMs) in generating human-aligned responses. However, collecting diverse, high-quality instruction data for tuning poses challenges, particularly in privacy-sensitive domains. Federated instruction tuning (FedIT) has emerged as a solution, leveraging federated learning from multiple data owners while preserving privacy. Yet, it faces challenges due to limited instruction data and vulnerabilities to training data extraction attacks. To address these issues, we propose a novel federated algorithm, FedPIT, which utilizes LLMs' in-context learning capability to self-generate task-specific synthetic data for training autonomously. Our method employs parameter-isolated training to maintain global parameters trained on synthetic data and local parameters trained on augmented local data, effectively thwarting data extraction attacks. Extensive experiments on real-world medical data demonstrate the effectiveness of FedPIT in improving federated few-shot performance while preserving privacy and robustness against data heterogeneity.

翻译：指令微调已被证明是提升大语言模型（LLMs）生成符合人类对齐响应的关键手段。然而，在隐私敏感领域中，收集多样化且高质量的指令微调数据仍面临挑战。联邦指令微调（FedIT）作为一种解决方案应运而生，它通过联邦学习实现多方数据持有者的协作，同时保护隐私。但该方法仍受限于指令数据不足以及易受训练数据提取攻击的问题。针对上述挑战，我们提出了一种新型联邦算法FedPIT，该算法利用LLMs的上下文学习能力自主生成任务特定的合成数据用于训练。该方法采用参数隔离训练策略，将基于合成数据训练的全局参数与基于增强本地数据训练的本地参数分离，有效抵御数据提取攻击。在真实医疗数据上的大量实验表明，FedPIT能有效提升联邦小样本场景下的性能，同时保持隐私性并对数据异构性具有鲁棒性。