Further raising awareness for the inherent vulnerabilities of Graph Neural Networks (GNNs) to adversarial Bit Flip Attacks (BFAs), Lorenz presented our group’s latest work “On the Relationship Between Robustness and Expressivity of Graph Neural Networks” by Lorenz Kummer, Wilfried Gansterer and Nils Kriege at the 28th International Conference on Artificial Intelligence and Statistics (AISTATS).

In our work, which extends our initial investigation of relationship between GNNs and BFAs from first GNN-specific attack and defense mechanisms presented previously at ACM SIGKDD 2024 and AAAI 2025, we offer a first theoretical analysis how architectural choices, graph properties, and feature encodings influence both the expressivity and vulnerability of GNNs. We find that ReLU-activated GNNs operating on highly homophilous graphs with low-dimensional features are particularly susceptible to BFAs.

Not only was Mai Khao, Thailand an incredible conference location, but the discussions we had also inspired novel research ideas that we intend to pursue in the near future.