Core Concepts
This work presents the first approach to formally verify graph convolutional networks with uncertain node features and uncertain graph structure.
Abstract
The key highlights and insights of this content are:
The authors present the first approach to formally verify graph convolutional networks with uncertain node features and uncertain graph structure as input.
The considered architecture of the graph convolutional network is generic and can have any element-wise activation function.
The approach allows verifying the graph neural network over multiple message-passing steps given an uncertain graph input.
The authors explicitly preserve the non-convex dependencies of all involved variables through all layers of the graph neural network using (matrix) polynomial zonotopes.
The verification algorithm has polynomial time complexity in the number of uncertain input features and in the number of uncertain edges.
The approach is demonstrated on three popular benchmark datasets with added perturbations on the node features and the graph structure.
The approach will be made publicly available with the next release of CORA.