Privacy Preserving Semi-supervised Learning for Labeled Graphs

We propose a novel privacy preserving learning algorithm that achieves semi-supervised learning in graphs. In real world networks, such as disease infection over individuals, links (contact) and labels (infection) are often highly sensitive information. Although traditional semisupervised learning methods play an important role in network data analysis, they fail to protect such sensitive information. Our solutions enable to predict labels of partially labeled graphs without disclosure of labels and links, by incorporating cryptographic techniques into the label propagation algorithm. Even when labels included in the graph are kept private, the accuracy of our PPLP is equivalent to that of label propagation which is allowed to observe all labels in the graph. Empirical analysis showed that our solution is scalable compared with existing privacy preserving methods. The results with human contact networks showed that our protocol takes only about 10 seconds for computation and no sensitive information is disclosed through the protocol execution.