With “Crossfire: An Elastic Defense Framework for Graph Neural Networks under Bit Flip Attacks” by Lorenz Kummer, Samir Moustafa, Wilfried Gansterer and Nils Kriege, we presented a novel approach to defending Graph Neural Networks (GNNs) from Bit Flip Attacks (BFAs) at the 39th Annual AAAI Conference on Artificial Intelligence.

Crossfire combines innovative techniques like saliency-based honeypot neuron induction, the exploitation of sparsity, and cryptographic hashing to detect and repair compromised weights efficiently on the bit-level. Defending against sophisticated, GNN-specific attacks such as the one introduced by our group in “Attacking Graph Neural Networks with Bit Flips: Weisfeiler and Leman Go Indifferent” at ACM SIGKDD 2024, it achieves near-perfect attack detection and restores models to their pre-attack state with significantly higher accuracy and reliability than existing defenses ported from the computer vision domain.

Overall, we were thrilled by the positive reception of our poster at AAAI and greatly appreciated the engaging discussions it sparked!