Adversarial Attacks on Graph Neural Networks via Meta Learning

By Daniel Zügner et al.

Published on Jan. 28, 2024

Read the original document by opening this link in a new tab.

Table of Contents

1 Introduction
2 Related Work
3 Problem Formulation
4 Graph Structure Poisoning via Meta-Learning

Summary

This paper investigates training time attacks on graph neural networks for node classification using meta-gradients to solve the bilevel problem. The attacks perturb the discrete graph structure to misguide the networks, leading to decreased performance. The authors propose a poisoning attack algorithm based on meta-learning, treating the graph as a hyperparameter to optimize. The study shows that small perturbations can significantly impact the performance of graph convolutional networks and unsupervised embeddings. The attacks do not require knowledge about the target classifiers and aim to compromise the global node classification performance.