Rethinking Molecular Graph Backdoors under Chemistry-aware Admission

The paper introduces ChemGuard, a protocol for assessing the admissibility of molecular records in graph neural networks (GNNs), addressing the inadequacies of existing defenses against backdoor attacks. It presents ChemBack, a novel admission-aware backdoor attack that leverages chemically feasible motifs and ranks them based on Tanimoto similarity to target molecules, demonstrating high success rates across various benchmarks while maintaining clean model accuracy. This highlights the need for practitioners to consider both admission protocols and the potential for sophisticated backdoor strategies that align with chemical validity in molecular GNNs.