Reverse Link Prediction Based on Core-periphery Detection

This work proposes a novel structure-aware algorithms framework for protecting sensitive links against attacks from link prediction. By leveraging the core-periphery structure commonly found in complex networks, we design a defense strategy using perturbations with a small impact on the original topology. Specifically, we employ a masked Bayesian matrix factorization algorithm to detect core-periphery structures and use a graph attention network to model hierarchical node representations. Based on node pair types, we apply differentiated perturbation strengths and optimize the framework via a multi-objective loss function that jointly considers prediction suppression and embedding separation. Extensive experiments on real-world networks demonstrate that our method significantly degrades link prediction performance while better preserving structural properties compared to baseline methods.