LinkGuard: Link Locally Privacy-Preserving Graph Neural Networks with Integrated Denoising and Private Learning

Recent studies have introduced privacy-preserving graph neural networks to safeguard the privacy of sensitive link information in graphs. However, existing link protection mechanisms in GNNs, particularly over decentralized nodes, struggle to strike an optimal balance between privacy and utility. We argue that a pivotal issue is the separation of noisy topology denoising and GNN private learning into distinct phases at the server side, leading to an under-denoising problem in the noisy topology. To address this, we propose a dynamic, adaptive Link LDP framework that performs noisy topology denoising on the server side in a dynamic manner. This approach aims to mitigate the impact of local noise on the GNN training process, reducing the uncertainty introduced by local noise. Furthermore, we integrate the noise generation and private training processes across all existing Link LDP GNNs into a unified framework. Experimental results demonstrate that our method surpasses existing approaches, obtaining around a 7% performance improvement under strong privacy strength and achieving a better trade-off between utility and privacy.