LLM-Driven Synthetic Text Generation for Privacy-Preserving Federated Learning

Federated learning (FL) enables collaborative fine tuning of pre-trained language models (PLMs) across user devices by aggregating locally trained model parameters rather than raw data, exhibiting great potential in privacy-sensitive natural language processing (NLP) applications. However, the inherent tendency of model parameters to memorize training data puts existing efforts at risk of disclosing raw data. Knowledge distillation (KD) offers a privacy-preserving alternative for knowledge transfer between the cloud server and user devices by exchanging lightweight model predictions associated with a public proxy dataset. Nevertheless, a desirable proxy dataset may not always be available in practice as its construction requires careful deliberation and even prior knowledge about the private domain to ensure satisfactory performance, which conflicts with the privacy-preserving principles of FL. To tackle this challenge, this paper proposes FedSTG, a privacy-preserving FL framework empowered by synthetic text generation of large language models (LLMs), which combines the exceptional generation capability of a pre-trained LLM deployed on the cloud server with the domain specific understanding capabilities of local PLMs distributed across user devices to generate high-quality synthetic text for knowledge transfer, facilitating the integration of KD and FL without the need for publicly available proxy data. Extensive experiments across a wide range of benchmark tasks demonstrate that FedSTG achieves superior performance without requiring any prior knowledge about the private domain, delivering a privacy-preserving FL solution for NLP.