Research map / Adversarial robustness and security / Black-box and transfer attacks
SPADE: A Spectral Method for Black-Box Adversarial Robustness Evaluation
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 |
|---|---|---|---|
| Use SPADE-guided variable perturbation sizes instead of random perturbation assignment | PGD adversarial training on MNIST · MNIST, epsilon 0.4 | 20PGD-100 accuracy: 83.09 -> 88.46 | Helped |
| Use SPADE-guided variable perturbation sizes instead of random perturbation assignment | PGD adversarial training on MNIST · MNIST, epsilon 0.4 | Clean accuracy: 97.81 -> 97.28 | Hurt |
| Use SPADE-guided variable perturbation sizes instead of random perturbation assignment | PGD adversarial training on CIFAR-10 · CIFAR-10, epsilon 2.0/4.0/8.0 | 10PGD-50 accuracy: 70.87/65.73/51.77 -> 75.74/69.19/53.61 | Mixed |
| Use SPADE-guided variable perturbation sizes instead of vanilla PGD | PGD adversarial training on CIFAR-10 · CIFAR-10, epsilon 2.0/4.0/8.0 | 10PGD-50 accuracy: 75.28/67.92/50.35 -> 75.74/69.19/53.61 | Helped |
| Use SPADE-guided variable perturbation sizes instead of vanilla PGD | PGD adversarial training on CIFAR-10 · CIFAR-10 | Clean accuracy: 81.57 -> 81.38 | Hurt |
| Choose edges using the largest edge SPADE scores rather than randomly | SPADE edge ranking · MNIST test set | Average DMD (10NN): 3.1/2.6/2.5/2.7 -> 6.6/6.1/6.5/7.7 | Helped |
| Choose edges using the largest edge SPADE scores rather than randomly | SPADE edge ranking · CIFAR-10 test set | Average DMD (10NN): 6.8/6.4/5.6/5.5/5.3 -> 11.4/11.5/9.7/12.5/8.0 | Helped |
| Increase kNN graph k from below 10 to 20 (ablation) | SPADE robustness ranking · MNIST | Correct model non-robustness ranking | Helped |
| Vary kNN graph k from 5 to 20 (ablation) | SPADE robustness ranking · CIFAR-10 | Correct model non-robustness ranking | No clear effect |
| Evaluate CLEVER using SPADE-selected vulnerable samples rather than random samples | CLEVER robustness evaluation · MNIST-2CL, epsilon 0 | CLEVER score: 162.35/7.592 -> 0.049/0.002 | Mixed |
| Use AugMix-trained models instead of standard-trained models | CIFAR-10 natural robustness evaluation · Clean test set | SPADE score: Allconv: 213.03 -> 149.63; Densenet: 152.07 -> 129.49; Resnext: 614.63 -> 571.15; WRN: 784.03 -> 738.35 | Helped |
| Use AugMix-trained models instead of standard-trained models | CIFAR-10 natural robustness evaluation · Corrupted test set | SPADE score: Allconv: 39.99 -> 60.68; Densenet: 33.63 -> 53.32; Resnext: 51.57 -> 293.94; WRN: 72.96 -> 100.17 | Hurt |
Closest papers by meaning
- A Spectral View of Adversarially Robust Features (NeurIPS 2018 · similarity 0.90)
- Adversarial Robustness of Supervised Sparse Coding (NeurIPS 2020 · similarity 0.87)
- Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks (NeurIPS 2025 · similarity 0.86)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.86)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.86)
- GREAT Score: Global Robustness Evaluation of Adversarial Perturbation using Generative Models (NeurIPS 2024 · similarity 0.86)
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.