Research map / Graph learning / GNN architectures
Efficient Robustness Certificates for Discrete Data: Sparsity-Aware Randomized Smoothing for Graphs, Images and More
Aleksandar Bojchevski, Johannes Gasteiger, Stephan Günnemann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use sparsity-aware data-dependent flip probabilities (p+=0.01, p-=0.6) instead of uniform flipping (p+=p-=0.1) (ablation) | GCN attribute certificates on Cora-ML · Cora-ML attribute perturbations | avg max certified radius rd: 2.03 -> 9.99 | Improved |
| Sparsity-aware smoothing with deletion-only flips (p+=0) instead of uniform flipping (ablation) | GCN attribute certificates on Cora-ML · Cora-ML attribute perturbations | avg max certified radius rd: 2.03 -> 18.66 | Improved |
| Sparsity-aware randomization in place of randomized ablation certificate (ablation) | GCN attribute certificates on Cora-ML · Cora-ML attribute perturbations | avg max certified radius rd: 2.01 -> 9.99 | Improved |
| Certificate with same flip probabilities as Lee et al. (2019) on discretized ImageNet, now computed via region partition | Certificate efficiency · Discretized ImageNet (K=256) | certificate computation time: 4 days -> 2.5 ms | Improved |
| Sparsity-aware flip probabilities (p+=0.1, p-=0.2) vs. uniform p+=p-=0.2 on binarized MNIST (ablation) | CNN certificates on binarized MNIST · Binarized MNIST | certified ratio w.r.t. rd | Improved |
| Multi-class (Bonferroni) certificate instead of binary-class certificate (ablation) | CNN certificates on binarized MNIST · Binarized MNIST | certified ratio | Improved |
| Multi-class certificate instead of binary-class certificate (ablation) | GCN on Cora-ML · Cora-ML attribute perturbations | certified ratio | Improved |
| Deletion-only smoothing (p+=0) instead of adding flips, certifying rd (ablation) | GAT, GCN, APPNP graph structure certificates · Cora-ML graph perturbations | certified ratio for rd | Improved |
| Certifying with GAT/GCN/APPNP models: attribute robustness ordering | GNN base models · Cora-ML attribute perturbations | certified ratio | Mixed |
| Joint smoothing of graph and attributes with separate flip probabilities/radii vs single-input certification | GCN joint certificates on Cora-ML · Cora-ML joint graph and attribute perturbations | certified ratio | Hurt |
Closest papers by meaning
- Hierarchical Randomized Smoothing (NeurIPS 2023 · similarity 0.91)
- Randomized Message-Interception Smoothing: Gray-box Certificates for Graph Neural Networks (NeurIPS 2022 · similarity 0.89)
- Incremental Randomized Smoothing Certification (ICLR 2024 · similarity 0.88)
- Certified Robustness for Deep Equilibrium Models via Serialized Random Smoothing (NeurIPS 2024 · similarity 0.88)
- Mitigating the Curse of Dimensionality for Certified Robustness via Dual Randomized Smoothing (ICLR 2024 · similarity 0.87)
- Robust Graph Condensation via Classification Complexity Mitigation (NeurIPS 2025 · 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.
