Illinois team receives NSF grant for safe graph neural networks

A research team led by the Illinois Grainger College of Engineering Siebel School of Computing and Data Science professor Hanghang Tong and Jingrui He, professor and MSIM program director in the School of Information Sciences and CS, CDS, NCSA and CDA affiliate, has received a new $800,000 NSF AI-Safety program grant for safe graph neural networks; “NetSafe: Towards a Computational Foundation of Safe Graph Neural Networks.”

Photo Credit: University of Illinois / Holly Birch Photography

Photo Credit: School of Information Sciences

Graph neural networks (GNNs) represent a family of deep learning methods designed for interrelated data. These data are called graph data because the graph is used to highlight the interconnections or nodes between the data points. GNNs produce representations of the nodes to explore the interrelationships and have provided a solution for a wide range of scientific applications, ranging from social media analysis to neuroscience, healthcare, climate science, finance, aviation, e-commerce, and biology. As the project abstract puts it, the fundamental safety issues of GNN have not been well studied, but “several important questions largely remain open.”

Tong says the project's investigative scope will cover “the three stages of GNNs, including training, adaptation, and testing” in its integral research thrusts. Tong defines the hazards associated with GNNs, saying, “In the context of GNN, we use ‘hazards’ to refer to (1) external perturbations on the training data used for training GNN, and (2) dataset shifts during the GNN adaptation and testing.”

Tong identifies one safety issue surrounding GNNs:

“Take financial fraud detection as an example. First, the financial transaction data is inherently noisy due to measurement errors, reporting limitations, and accounting errors (e.g., duplication, omission, or entry reversal), with inaccurate labels due to human errors such as unreported/undetected frauds and wrongly coded fraud types. These are the ‘external perturbations.’  Second, the behavior of fraudsters might change for the purpose of bypassing the current detector, and new types of fraud are likely to emerge over time. These are the ‘dataset shifts.’” He notes, "Professor He’s work on domain adaptation will address the dataset shifts during the adaptation and testing of GNN, which is a critical aspect of safe GNN.”

The research group will apply what they learn to coursework at the Siebel School of Computing and Data Science and the School of Information Sciences, Tong says. “The results of this project will be assimilated into the courses that we teach at Illinois, including CS412, CS512, CS514, and IS577, and IS407.” CS PhD students Zhichen Zeng, Ruizhong Qiu and Zhe Xu and iSchool PhD student Tianxin Wei round out the research team.