Stanford CS224W: ML with Graphs | 2021 | Lecture 16.4 - Robustness of Graph Neural Networks

For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: stanford.io/3Csn5T7
Jure Leskovec
Computer Science, PhD
For the last segment of our discussion on advanced GNN topics, we discuss the robustness of GNNs. We first introduce the fact that deep learning models are vulnerable to adversarial attacks. We then discuss the possibilities of adversarial attacks over graphs, including direct and indirect attacks. We then review the GCN model that we are going to attack, where we mathematically formalize the attack problem as an optimization problem. Finally, we empirically see how vulnerable GCN’s predictions are to the adversarial attacks.
To follow along with the course schedule and syllabus, visit:
web.stanford.edu/class/cs224w/
#machinelearning #machinelearningcourse