# Image Classification on MNIST (test) (Adversarial Robustness: PGD, CW, AA)

Source: [Image Classification on MNIST (test) (Adversarial Robustness: PGD, CW, AA)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T05:34:20.832Z

Evaluation of Image Classification performance on the MNIST test set, focusing on standard accuracy and robustness against specific adversarial attacks (PGD, CW, AA).

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Image Classification |
| Dataset (including any stated split) | MNIST (test) |
| Metric | Accuracy (Natural) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 15 of 15 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Accuracy (Natural)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa.md?metric=Accuracy%20(Natural)>): Higher is better; unit: PERCENT.

- [Accuracy (PGD)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa.md?metric=Accuracy%20(PGD)>): Higher is better; unit: PERCENT.

- [Accuracy (CW)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa.md?metric=Accuracy%20(CW)>): Higher is better; unit: PERCENT.

- [Accuracy (AA)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa.md?metric=Accuracy%20(AA)>): Higher is better; unit: PERCENT.

- [Accuracy (PGD-40)](<https://www.sota2.com/research/sota/image-classification-on-mnist-test-adversarial-robustness-pgd-cw-aa.md?metric=Accuracy%20(PGD-40)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Accuracy (Natural) | Result date | Report paper |
| --- | --- | --- | --- |
| Oracle | 99.46 | Not recorded | 68fa6dc41e209131cea3bff4 |
| Warm-up+PLA | 99.22 | Not recorded | 68fa6dc41e209131cea3bff4 |
| Two-stage | 99.07 | Not recorded | 68fa6dc41e209131cea3bff4 |
| Clean Training | 98.98 | Not recorded | 69520ae153fdd5c36173d43f |
| Selective Adversarial Training (Grad Matching) | 98.26 | Not recorded | 69520ae153fdd5c36173d43f |
| Selective Adversarial Training (Margin-based) | 98.11 | Not recorded | 69520ae153fdd5c36173d43f |
| Full PGD-AT | 97.78 | Not recorded | 69520ae153fdd5c36173d43f |
| Random Subset PGD | 97.37 | Not recorded | 69520ae153fdd5c36173d43f |
| FORWARD | 97.22 | Not recorded | 68fa6dc41e209131cea3bff4 |
| LOG | 97.16 | Not recorded | 68fa6dc41e209131cea3bff4 |
| SCL\_NL | 93.09 | Not recorded | 68fa6dc41e209131cea3bff4 |
| NN | 68.48 | Not recorded | 68fa6dc41e209131cea3bff4 |
| FREE | 48.94 | Not recorded | 68fa6dc41e209131cea3bff4 |
| SCL\_EXP | 14.88 | Not recorded | 68fa6dc41e209131cea3bff4 |
| EXP | 10.99 | Not recorded | 68fa6dc41e209131cea3bff4 |

### Oracle

Accuracy (Natural): 99.46. Result ID: 698fdb273a6c314849fc5b6e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa665a1e209131cea0fbae.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Warm-up+PLA

Accuracy (Natural): 99.22. Result ID: 698fdb273a6c314849fc5b77.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Implementation](<https://github.com/RoyalSkye/ATCL>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Two-stage

Accuracy (Natural): 99.07. Result ID: 698fdb273a6c314849fc5b6f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Clean Training

Accuracy (Natural): 98.98. Result ID: 69993b285698665ec4a4f61a.

[Reporting paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

Recorded evaluation configuration: {"Adversarial Perturbation":"None"}. This record does not assert independent reproduction.

### Selective Adversarial Training (Grad Matching)

Accuracy (Natural): 98.26. Result ID: 69993b285698665ec4a4f61e.

[Reporting paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

Recorded evaluation configuration: {"Sampling Strategy":"Grad Matching","ρ (subset ratio)":0.25}. This record does not assert independent reproduction.

### Selective Adversarial Training (Margin-based)

Accuracy (Natural): 98.11. Result ID: 69993b285698665ec4a4f61d.

[Reporting paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

Recorded evaluation configuration: {"Sampling Strategy":"Margin-based","ρ (subset ratio)":0.25}. This record does not assert independent reproduction.

### Full PGD-AT

Accuracy (Natural): 97.78. Result ID: 69993b285698665ec4a4f61c.

[Reporting paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa665a1e209131cea0fbae.md>)

Recorded evaluation configuration: {"Sampling Strategy":"Full Minibatch","Perturbation":"Multi-step PGD"}. This record does not assert independent reproduction.

### Random Subset PGD

Accuracy (Natural): 97.37. Result ID: 69993b285698665ec4a4f61b.

[Reporting paper](<https://www.sota2.com/research/paper/69520ae153fdd5c36173d43f.md>)

Recorded evaluation configuration: {"Sampling Strategy":"Uniform Random","ρ (subset ratio)":0.25}. This record does not assert independent reproduction.

### FORWARD

Accuracy (Natural): 97.22. Result ID: 698fdb273a6c314849fc5b70.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64a41e209131cea05cc7.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### LOG

Accuracy (Natural): 97.16. Result ID: 698fdb273a6c314849fc5b76.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d351e209131cea386a5.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### SCL\_NL

Accuracy (Natural): 93.09. Result ID: 698fdb273a6c314849fc5b73.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68351e209131cea1acc9.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### NN

Accuracy (Natural): 68.48. Result ID: 698fdb273a6c314849fc5b72.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66ab1e209131cea1186d.md>)

[Implementation](<https://github.com/takashiishida/comp>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### FREE

Accuracy (Natural): 48.94. Result ID: 698fdb273a6c314849fc5b71.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66ab1e209131cea1186d.md>)

[Implementation](<https://github.com/takashiishida/comp>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### SCL\_EXP

Accuracy (Natural): 14.88. Result ID: 698fdb273a6c314849fc5b74.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68351e209131cea1acc9.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### EXP

Accuracy (Natural): 10.99. Result ID: 698fdb273a6c314849fc5b75.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6dc41e209131cea3bff4.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d351e209131cea386a5.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
