Research map / Graph learning / GNN architectures
Rethinking Molecular Graph Backdoors under Chemistry-aware Admission
Thinh Nguyen, Sze Jue Yang, Khoa D Doan, Chee Seng Chan, Kok-Seng Wong
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ChemGuard, a chemistry-aware admission protocol (sanitization plus graph-string consistency) for molecular backdoor evaluation, and ChemBack, a model-free motif-attachment backdoor that selects admissible triggers by Tanimoto similarity to target-class molecules and stays effective under admission.
Problem
Molecular GNN backdoor attacks are usually evaluated as abstract graph edits, ignoring that real molecular pipelines require records to pass parsing, sanitization, canonicalization, and graph-string consistency checks before training or inference.
Approach
Introduces an admission-stage protocol (ChemGuard) that counts only chemically sanitizable, graph-string-consistent poisons and test triggers, and a black-box attack (ChemBack) that searches valid motif-anchor attachments ranked by fingerprint Tanimoto similarity to clean target-class molecules, without victim or surrogate model access.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enforce ChemGuard admission | MB · BACE (10% poison rate) | ASR (%): 82.67 -> 13.38 | Hurt |
| enforce ChemGuard admission | UGBA · Tox21 (10% poison rate) | ASR (%): 97.06 -> 22.84 | Hurt |
| enforce ChemGuard admission | GTA · BBBP (10% poison rate) | ASR (%): 73.16 -> 54.47 | Hurt |
| enforce ChemGuard admission | ChemBack · BBBP (10% poison rate) | ASR (%): 69.36 -> 69.36 | No clear effect |
| replace graph-only attack (MB) with ChemBack under ChemGuard | MB · Tox21 (with ChemGuard, 10% poison rate) | ASR (%): 24.61 -> 81.84 | Improved |
| add Tanimoto target alignment (under Spectral Signatures defense) (ablation) | ChemBack · BACE (Spectral Signatures, 10% poison rate) | ASR (%): 42.36 -> 97.84 | Improved |
| add Tanimoto target alignment (under Spectral Signatures defense) (ablation) | ChemBack · BBBP (Spectral Signatures, 10% poison rate) | ASR (%): 31.48 -> 68.74 | Improved |
| add Tanimoto target alignment (no defense) (ablation) | ChemBack · BBBP (with ChemGuard, 10% poison rate) | ASR (%): 68.91 -> 69.36 | No clear effect |
| add validity pre-screening (ablation) | ChemBack · BBBP (with ChemGuard, 10% poison rate) | ASR (%): 69.02 -> 69.36 | No clear effect |
| add validity pre-screening (under Spectral Signatures defense) (ablation) | ChemBack · BBBP (Spectral Signatures, 10% poison rate) | ASR (%): 67.06 -> 68.74 | No clear effect |
| apply RIGBD defense | ChemBack · SIDER (10% poison rate) | ASR (%): 99.17 -> 92.74 | Hurt |
| apply RIGBD defense | ChemBack · BBBP (10% poison rate) | ASR (%): 69.36 -> 61.37 | Hurt |
Surprising findings
- Graph-only baselines lose much of their apparent ASR once chemically invalid poisons and triggers are rejected.
Limitations
- ChemGuard is a first-layer admission check and not a complete defense or robustness certificate; it does not cover synthetic accessibility, medicinal-chemistry rules, or expert review.
- Trigger space limited to single-step motif attachments.
- Tanimoto alignment and MD2 diagnostics do not guarantee universal stealth.
Experimental setup
Datasets: BBBP BACE SIDER Tox21
Benchmarks: BBBP BACE SIDER Tox21
Metrics: clean accuracy attack success rate effective poisoning rate
Baselines: GTA MB UGBA DPGBA
Builds on: GTA UGBA DPGBA Motif-Backdoor
Closest papers by meaning
- Extracting Training Data from Molecular Pre-trained Models (NeurIPS 2024 · similarity 0.87)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.85)
- Molecular Data Programming: Towards Molecule Pseudo-labeling with Systematic Weak Supervision (CVPR 2024 · similarity 0.85)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.85)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.85)
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
