Quantifying Privacy Leakage in Graph Embedding

Graph embedding, Graph embedding and privacy

This paper quantifies the privacy leakage in graph embedding with three attacks: - membership inference: whether a node was a member of the model’s training or not - graph reconstruction: how well can we reconstruct the graph using the embedding? - attribute inference: inferring sensitive attributes using graph embedding.

“Blackbox” setting gives only access to the model outputs for a given input x, not the embedding itself. “Whitebox” setting provides the embedding as well.