Research map / Adversarial robustness and security / Certified and empirical robustness
Generalization of Neural Combinatorial Solvers Through the Lens of Adversarial Robustness
Simon Geisler, Johanna Sommer, Jan Schuchardt, Aleksandar Bojchevski, 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 |
|---|---|---|---|
| Adversarially fine-tune NeuroSAT for 10 extra epochs using perturbed SAT instances | NeuroSAT · 10-40 Attack | accuracy: 50.0 ± 1.16 -> 54.0 ± 0.48 | Improved |
| Adversarially fine-tune NeuroSAT for 10 extra epochs using perturbed SAT instances | NeuroSAT · 50-100 Attack | accuracy: 39.4 ± 3.15 -> 44.4 ± 1.19 | Improved |
| Adversarially fine-tune NeuroSAT for 10 extra epochs using perturbed SAT instances | NeuroSAT · 100-300 | accuracy: 64.4 ± 1.53 -> 67.0 ± 0.74 | Improved |
| Adversarially fine-tune NeuroSAT for 10 extra epochs using perturbed SAT instances | NeuroSAT · SATLIB | accuracy: 66.1 ± 3.07 -> 63.7 ± 1.88 | Hurt |
| Train NeuroSAT on a 10% larger training set instead of regular training | NeuroSAT · 10-40 Attack | accuracy: 50.0 ± 1.16 -> 49.6 ± 1.52 | No clear effect |
| Attack NeuroSAT with sound SAT-preserving literal perturbations | NeuroSAT · 10-40 satisfiable instances | accuracy: 26% | Hurt |
| Use a SAT perturbation budget of 0.2% of literals | NeuroSAT · SAT satisfiable instances | recall: 0.2% | Hurt |
| Add five sound adversarial nodes to DecisionTSP instances | DecisionTSP · TSP with n ∼ U(20,40) nodes | accuracy | Hurt |
| Add five sound adversarial nodes to ConvTSP instances | ConvTSP · TSP with n = 20 nodes | optimality gap: 2% | Hurt |
| Adversarially fine-tune NeuroSAT with a 5% budget on satisfiable instances (ablation) | NeuroSAT · SAT satisfiable instances | accuracy: 5% | Hurt |
| Use adversarial fine-tuning rather than regular training | NeuroSAT · 3-10 | accuracy: 92.3 ± 0.57 -> 93.0 ± 0.33 | Improved |
Closest papers by meaning
- ROCO: A General Framework for Evaluating Robustness of Combinatorial Optimization Solvers on Graphs (ICLR 2023 · similarity 0.89)
- Formalizing Generalization and Adversarial Robustness of Neural Networks to Weight Perturbations (NeurIPS 2021 · similarity 0.88)
- Learning perturbation sets for robust machine learning (ICLR 2021 · similarity 0.88)
- Sensitivity and Generalization in Neural Networks: an Empirical Study (ICLR 2018 · similarity 0.88)
- Adversarially Robust Learning: A Generic Minimax Optimal Learner and Characterization (NeurIPS 2022 · similarity 0.87)
- Why Robust Generalization in Deep Learning is Difficult: Perspective of Expressive Power (NeurIPS 2022 · 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.
