Image Classification on MNIST (Adversarial Robustness Metrics)
94.5Clean AccuracyVanilla
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VanillaAttack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.5 | 85.8 | 92.1 | 56,000 | |
| AdvTrain (off)Attack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.5 | 81.1 | 89.2 | 103,000 | |
| Selective (off)Attack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.5 | 88.7 | 93.3 | 94,000 | |
| Selective (on)Attack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.5 | 89.2 | 93.5 | 95,000 | |
| Two-Stage Sel.Attack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.5 | 89.4 | 93.7 | 89,000 | |
| Sep-AwareAttack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.2 | 78 | 87.5 | 111,000 | |
| AdvTrain (on)Attack=Black-Box Transfer PGD-20, ρ=0.9, Surrogate Model=SimpleCNN2026.05 | 94.1 | 79.4 | 88.2 | 111,000 |