[Paper Review] Federated Domain-Specific Knowledge Transfer on Large Language Models Using Synthetic Data

Haoran Li, Xinyuan Zhao+7 more|arXiv (Cornell University)|May 23, 2024

Topic Modeling4 citations

TL;DR

This paper proposes Federated Domain-Specific Knowledge Transfer (FDKT), a privacy-preserving framework that uses differentially private synthetic data from clients’ private domain data to enable large language models (LLMs) to transfer domain-specific knowledge to small language models (SLMs) without exposing raw data. FDKT achieves up to 5% accuracy gain in SLMs across diverse domains while maintaining a privacy budget of less than 10.

ABSTRACT

As large language models (LLMs) demonstrate unparalleled performance and generalization ability, LLMs are widely used and integrated into various applications. When it comes to sensitive domains, as commonly described in federated learning scenarios, directly using external LLMs on private data is strictly prohibited by stringent data security and privacy regulations. For local clients, the utilization of LLMs to improve the domain-specific small language models (SLMs), characterized by limited computational resources and domain-specific data, has attracted considerable research attention. By observing that LLMs can empower domain-specific SLMs, existing methods predominantly concentrate on leveraging the public data or LLMs to generate more data to transfer knowledge from LLMs to SLMs. However, due to the discrepancies between LLMs' generated data and clients' domain-specific data, these methods cannot yield substantial improvements in the domain-specific tasks. In this paper, we introduce a Federated Domain-specific Knowledge Transfer (FDKT) framework, which enables domain-specific knowledge transfer from LLMs to SLMs while preserving clients' data privacy. The core insight is to leverage LLMs to augment data based on domain-specific few-shot demonstrations, which are synthesized from private domain data using differential privacy. Such synthetic samples share similar data distribution with clients' private data and allow the server LLM to generate particular knowledge to improve clients' SLMs. The extensive experimental results demonstrate that the proposed FDKT framework consistently and greatly improves SLMs' task performance by around 5\% with a privacy budget of less than 10, compared to local training on private data.

Motivation & Objective

• To address the challenge of improving small language models (SLMs) in data-scarce, privacy-sensitive domains without exposing private data.
• To overcome the distribution mismatch between public data and clients’ private domain data in existing knowledge distillation methods.
• To enable effective knowledge transfer from server-side LLMs to client-side SLMs while preserving data privacy and model intellectual property.
• To develop a scalable and versatile framework applicable across diverse model architectures and domain-specific tasks.

Proposed method

• Clients generate differentially private synthetic data from their private domain data using a private generator, ensuring privacy under a bounded privacy budget.
• The synthetic data is transmitted to the server, where a large language model (LLM) performs clustering-based filtering to detect and remove artifacts introduced by differential privacy.
• The cleaned synthetic data is used to condition in-context learning on the server-side LLM, enabling targeted data augmentation that reflects the client’s domain distribution.
• The augmented data is sent back to the client for fine-tuning the local SLM, enabling selective knowledge transfer without exposing raw data.
• The framework supports API-level access to the server LLM, protecting both client data and server model weights.
• The method is agnostic to model architectures, enabling broad applicability across various LLMs and SLMs.

Experimental results

Research questions

• RQ1 Can synthetic data generated under differential privacy preserve domain-specific distributional characteristics while ensuring client data privacy?
• RQ2 How effective is the proposed filtering mechanism in mitigating artifacts introduced by differential privacy in synthetic data?
• RQ3 To what extent can knowledge transfer from a server-side LLM improve the performance of client-side SLMs on low-resource, domain-specific tasks?
• RQ4 How does FDKT perform across multiple clients with diverse domain-specific tasks in a multi-client, multi-task setting?

Key findings

• FDKT consistently improves SLM performance by approximately 5% across multiple domains, including Yelp Health, Shopping, and AGNews, even with as few as 200 private training samples.
• In the one-to-many scenario, FDKT significantly reduces the performance gap between in-domain and out-of-domain SLM evaluations, demonstrating improved multi-task generalization.
• The framework maintains strong privacy guarantees, with a privacy budget of less than 10, while still enabling effective knowledge transfer.
• Performance gains are highly correlated with the capability of the server-side LLM, with stronger LLMs yielding better results across all evaluated models, including Mistral-7b, Llama2-7b, Llama3-8b, and Qwen-14b.
• Ablation studies confirm that both the privacy budget and the number of augmented samples significantly influence performance, with optimal gains observed at moderate settings.