Privately Learning from Graphs with Applications in Fine-tuning Large Language Models

2024-10-10

arXiv.org

AI Summary

Conventional differential privacy (DP) methods—such as DP-SGD—fail on graph-structured data due to inherent structural dependencies among nodes, undermining privacy guarantees during fine-tuning of large language models (LLMs) on text-attributed graphs. Method: This paper proposes the first DP-compliant fine-tuning framework tailored to graph structural dependencies. It introduces a graph-relational decoupling sampling mechanism to relax the i.i.d. sample assumption of DP-SGD; designs a privacy-preserving fine-tuning paradigm integrating GNN-based graph embedding with LLM adaptation (compatible with BERT and Llama2); and establishes a tripartite trade-off analysis framework balancing privacy (ε), utility, and efficiency. Results: Evaluated on four real-world text-attributed graphs, the framework achieves an average 12.7% improvement in F1 score while strictly satisfying ε ≤ 4 DP guarantees. The implementation is open-sourced for full reproducibility.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Natural Language Processing: Safety and Robustness
• Data Mining & Knowledge Management: Graph Mining, Social Network Analysis & Community

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Graphs offer unique insights into relationships and interactions between entities, complementing data modalities like text, images, and videos. By incorporating relational information from graph data, AI models can extend their capabilities beyond traditional tasks. However, relational data in sensitive domains such as finance and healthcare often contain private information, making privacy preservation crucial. Existing privacy-preserving methods, such as DP-SGD, which rely on gradient decoupling assumptions, are not well-suited for relational learning due to the inherent dependencies between coupled training samples. To address this challenge, we propose a privacy-preserving relational learning pipeline that decouples dependencies in sampled relations during training, ensuring differential privacy through a tailored application of DP-SGD. We apply this method to fine-tune large language models (LLMs) on sensitive graph data, and tackle the associated computational complexities. Our approach is evaluated on LLMs of varying sizes (e.g., BERT, Llama2) using real-world relational data from four text-attributed graphs. The results demonstrate significant improvements in relational learning tasks, all while maintaining robust privacy guarantees during training. Additionally, we explore the trade-offs between privacy, utility, and computational efficiency, offering insights into the practical deployment of our approach. Code is available at https://github.com/Graph-COM/PvGaLM.

Problem

• Learning from sensitive graph data with privacy
• Incompatibility of DP-SGD with relational learning
• Fine-tuning LLMs on private graph-structured information

Innovation

• Decouples dependencies in sampled relations
• Applies tailored DP-SGD for differential privacy
• Enables LLM fine-tuning on sensitive graph data