Measuring the Privacy Leakage via Graph Reconstruction Attacks on Simplicial Neural Networks (Student Abstract)

In this paper, we measure the privacy leakage via studying whether graph representations can be inverted to recover used generate them reconstruction attack (GRA). We propose a GRA that recovers graph's adjacency matrix from decoder minimizes loss between partial and reconstructed graph. study three types of are trained on graph, i.e., output convolutional network (GCN), attention (GAT), our proposed simplicial neural (SNN) higher-order combinatorial Laplacian. Unlike first two only encode pairwise relationships, third type representation, SNN outputs, encodes interactions (e.g., homological features) nodes. find outputs reveal lowest privacy-preserving ability defend GRA, followed by those GATs GCNs, which indicates importance building more private with node information could potential threats, such as GRAs.