Research map / Adversarial robustness and security / Certified and empirical robustness
Provable Adversarial Robustness for Group Equivariant Tasks: Graphs, Point Clouds, Molecules, and More
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase Gaussian smoothing standard deviation from σ=0.05 to σ=0.25 (ablation) | Smoothed PointNet on ModelNet40 · ModelNet40 | natural accuracy: 88% -> 75.5% | Hurt |
| Increase Gaussian smoothing standard deviation from σ=0.05 to σ=0.25 (ablation) | Smoothed DGCNN on ModelNet40 · ModelNet40 | natural accuracy: 90.4% -> 79.9% | Hurt |
| Use DGCNN instead of PointNet as the smoothed base model (ablation) | Point cloud classification · ModelNet40 | certified accuracy: up to 6.3 p.p. | Helped |
| Increase attribute insertion cost from c_X+=1 to c_X+=4, holding c_X-=1 (ablation) | Convex outer adversarial polytope certificate for a 2-layer GCN · Citeseer | certified accuracy at ε=25: 6.6% -> 23.4% | Helped |
| Increase attribute deletion cost while holding the other cost at 1 (ablation) | Convex outer adversarial polytope certificate for GCNs · Cora-ML and Citeseer | certified accuracy | No clear effect |
| Vary a single attribute insertion or deletion cost while holding the other at 1 (ablation) | Interval bound propagation certificate for GCNs · Cora-ML and Citeseer | certified accuracy | No clear effect |
| Increase edge insertion cost to c_A+=4 (ablation) | Policy iteration certificate for π-PPNP · Cora-ML and Citeseer | certified accuracy: almost tripling | Helped |
| Increase edge deletion cost c_A- from 1 to 4 (ablation) | Policy iteration certificate for π-PPNP · Citeseer | certified accuracy: up to 2.9 p.p. | Helped |
| Use edge deletion cost 4 and insertion cost 1 rather than insertion cost 4 and deletion cost 1 (ablation) | Linearization and dualization certificate for a GCN graph classifier · PROTEINS | average certified accuracy: 55.6% -> 61.2% | Helped |
| Use SphereNet instead of SchNet as the base model (ablation) | Smoothed molecular force prediction · MD17 | average test error: 0.51 to 1.89 -> 0.45 to 1.4 | Helped |
| Use SphereNet instead of SchNet as the base model (ablation) | Smoothed molecular force prediction · MD17 | certified output distance δ: 0.04 to 0.12 -> 0.35 to 1.61 | Hurt |
| Increase smoothing standard deviation σ (ablation) | Equivariance-preserving Gaussian smoothing for molecular force prediction · MD17 | maximum certifiable registration distance: grows linearly with σ | Helped |
Closest papers by meaning
- Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness (NeurIPS 2025 · similarity 0.91)
- A Spectral View of Adversarially Robust Features (NeurIPS 2018 · similarity 0.89)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.89)
- Can we have it all? On the Trade-off between Spatial and Adversarial Robustness of Neural Networks (NeurIPS 2021 · similarity 0.89)
- Exploring the Landscape of Spatial Robustness (ICML 2019 · similarity 0.88)
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.87)
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.