New protocol ChemGuard reveals vulnerabilities in molecular GNN backdoor attacks

Researchers have introduced ChemGuard, a new protocol to evaluate backdoor attacks on molecular graph neural networks (GNNs) by considering the chemical validity and consistency of molecular data. This approach reveals that many existing graph-based backdoor attacks lose effectiveness when subjected to realistic admission checks. To address this, a new attack method called ChemBack was developed, which constructs chemically feasible poisons that pass admission and are similar to clean molecules, demonstrating a significant threat even with these new defenses. AI

IMPACT Introduces new methods for assessing the security of molecular graph neural networks against sophisticated backdoor attacks.

RANK_REASON Academic paper detailing a new method and protocol for evaluating security vulnerabilities in machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• ChemBack
• ChemGuard
• DagsHub
• Gotit.pub
• graph neural network
• Hugging Face
• IArxiv
• ScienceCast
• Tanimoto similarity

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →