Robustness Certification on MNIST (test)
86.55Generalization BoundPartition-based (S4)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Partition-based (S4)Bound (Theorem)=S4, Approach type=Baselines, Target Model=CNN2026.02 | 86.55 | — | |
| Best Coreset (S9)Bound (Theorem)=S9, Approach type=Our approach, Target Model=CNN2026.02 | 21.5 | — | |
| PAC-Bayes (S15)Bound (Theorem)=S15, Approach type=Our approach, Target Model=CNN2026.02 | 4.83 | — | |
| Pick-To-Learn (S10)Bound (Theorem)=S10, Approach type=Our approach, Target Model=CNN2026.02 | 3.9 | — | |
| Model compression (S11)Bound (Theorem)=S11, Approach type=Our approach, Target Model=CNN2026.02 | 3.45 | — | |
| Norm-based (S5)Bound (Theorem)=S5, Approach type=Baselines, Target Model=CNN2026.02 | 3.14 | — | |
| Random Coreset (S8)Bound (Theorem)=S8, Approach type=Our approach, Target Model=CNN2026.02 | 1.79 | — | |
| ERANAbstract Domain=DeepPoly, Perturbation Radius (epsilon_linf)=0.011, Number of test points=1002026.03 | — | 99 | |
| FP (Ours)Arithmetic Assumption=Floating-Point, Perturbation Radius (epsilon_l2)=0.3, Number of test points=10,000, Method Variant=Hybrid pre-deployment method2026.03 | — | 95.27 | |
| Real (Ours)Arithmetic Assumption=Real, Perturbation Radius (epsilon_l2)=0.3, Number of test points=10,0002026.03 | — | 95.4 |