Community Concealment from Unsupervised Graph Learning-Based Clustering

A new research paper introduces a method to conceal sensitive communities from Graph Neural Network (GNN) clustering, a common AI surveillance technique. The defensive strategy makes limited, utility-aware changes by rewiring edges and modifying node features to reduce a community's distinctiveness. It outperforms the baseline DICE method, achieving median relative concealment improvements of 20-45% across synthetic and real network tests. This highlights a critical group-level privacy risk in graph learning.