Privacy-Preserving Deep Learning Over Graphs

Information Processing and Communications Lab, Imperial College London, London, UK


Graphs are a prevalent and ubiquitous kind of data structure that can represent rich relationship information between different entities. Recommending friendship in social networks, predicting protein-protein interactions, learning molecular fingerprints, and classifying diseases are kinds of tasks that are made possible by learning over graph data. In the past few years, the machine learning research over graphs has been revolutionized with the emergence of graph-based deep learning models, called graph neural networks (GNNs), which have demonstrated superior performance in the automatic learning of graph representation for various downstream tasks. However, learning over graphs can raise privacy concerns when they represent sensitive interactions or contain personal information. Previous works on privacy-preserving machine learning have presented effective solutions to protect the privacy of users when dealing with euclidean data, such as image, audio, and text, but addressing the privacy issues involved in applying deep learning algorithms on graphs is often more challenging due to the existing connectivity between graph nodes. This talk aims to give a gentle introduction to graph neural networks and the potential privacy risks of applying GNNs over sensitive graphs. Then, an overview of possible privacy attacks on GNNs and the most recent attempts to build privacy-preserving GNNs will be presented. The talk will conclude with a short discussion of key open problems and emerging research directions.

Download the slides