Is There a Structural Privacy Risk in Graph Prompting with LLMs?

Graph Prompting with Large Language Models (LLMs) can effectively integrate Knowledge Graphs (KGs) into downstream tasks of pre-trained LLMs. However, whether external KGs face a potential privacy risk for sensitive entity relationships during graph prompts training remains a pending issue. In this paper, we successfully conduct a Link Inference Privacy (LIP) attack on graph prompts, achieving a privacy attack success rate of over 80%. To address this issue, we propose a Structural Differentially Private Graph Neural Prompting (SDPGNP). Specifically, to mitigate privacy risks in the higher-order interactions of Graph Neural Networks (GNNs), we employ private statistics to obscure the existence of individual edges (edge-level privacy) in DP-Aggregation. Additionally, we analyze privacy risks related to graph structure relations in cross-modal interactions and propose DP-Alignment to enhance privacy through perturbation of cross-modal attention weights. Experiments on the OBQA, CSQA, and Riddle datasets demonstrate that compared with PromptDPSGD, SDPGNP maintains more stable training convergence speed and model performance under the same privacy budget, and significantly improves the resistance to LIP attacks by virtue of the effective integration of GNN’s robustness and noise mechanism in the specific scenarios of GNP. Notably, on the OBQA dataset, the attack rate decreases substantially from 94.80% to 49.95%.