# Adversarial Attack Perturbation Analysis on ImageNet

Source: [Adversarial Attack Perturbation Analysis on ImageNet](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet>)

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

Data updated: 2026-02-26T05:57:18.151Z

Evaluation of adversarial attacks on the ImageNet dataset, focusing specifically on the characteristics and quality of the generated perturbations using metrics like Parsimon, Square, SignH, and Superpixel counts.

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 | Adversarial Attack |
| Dataset (including any stated split) | ImageNet |
| Metric | Parsimon |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 19 of 19 |
| 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

- [Parsimon](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=Parsimon>): Higher is better; unit: PERCENT.

- [Square](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=Square>): Higher is better; unit: PERCENT.

- [SignH](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=SignH>): Higher is better; unit: PERCENT.

- [Acc (SignH)](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=Acc%20(SignH)>): Higher is better; unit: PERCENT.

- [Superpixel](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=Superpixel>): Higher is better; unit: PERCENT.

- [Perturbation Difference](<https://www.sota2.com/research/sota/adversarial-attack-perturbation-analysis-on-imagenet.md?metric=Perturbation%20Difference>): Lower is better; unit: SCALAR.

## Recorded results

| Method | Parsimon | Result date | Report paper |
| --- | --- | --- | --- |
| PyTorch ResNet-50 | 72.04 | Not recorded | 692fc3e7d6b0dd334293318c |
| Salman ResNet-18 | 58.6 | Not recorded | 692fc3e7d6b0dd334293318c |
| Wong ResNet-50 | 56.62 | Not recorded | 692fc3e7d6b0dd334293318c |
| Engstrom ResNet-50 | 48.92 | Not recorded | 692fc3e7d6b0dd334293318c |
| Salman ResNet-50 | 46.96 | Not recorded | 692fc3e7d6b0dd334293318c |
| Salman WideResNet-50-2 | 42.94 | Not recorded | 692fc3e7d6b0dd334293318c |
| Debenedetti XCiT-S12 | 37.44 | Not recorded | 692fc3e7d6b0dd334293318c |
| Debenedetti XCiT-M12 | 36.04 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ViT-S+ConvStem | 35.5 | Not recorded | 692fc3e7d6b0dd334293318c |
| Debenedetti XCiT-L12 | 35.32 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ConvNeXt-T+ConvStem | 34.88 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ConvNeXt-S+ConvStem | 33.36 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ViT-B+ConvStem | 31.02 | Not recorded | 692fc3e7d6b0dd334293318c |
| Liu Swin-B | 30.88 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ConvNeXt-B+ConvStem | 30.78 | Not recorded | 692fc3e7d6b0dd334293318c |
| Liu ConvNeXt-B | 30.32 | Not recorded | 692fc3e7d6b0dd334293318c |
| Singh ConvNeXt-L+ConvStem | 29.24 | Not recorded | 692fc3e7d6b0dd334293318c |
| Liu ConvNeXt-L | 28.92 | Not recorded | 692fc3e7d6b0dd334293318c |
| Liu Swin-L | 28.18 | Not recorded | 692fc3e7d6b0dd334293318c |

### PyTorch ResNet-50

Parsimon: 72.04. Result ID: 69970b7d5698665ec458ef4c.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

Recorded evaluation configuration: {"source":"PyTorch","Architecture":"ResNet-50","Iterations":"1000"}. This record does not assert independent reproduction.

### Salman ResNet-18

Parsimon: 58.6. Result ID: 69970b7d5698665ec458ef4a.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68e21e209131cea1ef57.md>)

[Implementation](<https://github.com/Microsoft/robust-models-transfer>)

Recorded evaluation configuration: {"source":"Salman","Architecture":"ResNet-18","Iterations":"1000"}. This record does not assert independent reproduction.

### Wong ResNet-50

Parsimon: 56.62. Result ID: 69970b7d5698665ec458ef47.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66e51e209131cea12efc.md>)

Recorded evaluation configuration: {"source":"Wong","Architecture":"ResNet-50","Iterations":"1000"}. This record does not assert independent reproduction.

### Engstrom ResNet-50

Parsimon: 48.92. Result ID: 69970b7d5698665ec458ef48.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

Recorded evaluation configuration: {"source":"Engstrom","Architecture":"ResNet-50","Iterations":"1000"}. This record does not assert independent reproduction.

### Salman ResNet-50

Parsimon: 46.96. Result ID: 69970b7d5698665ec458ef49.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68e21e209131cea1ef57.md>)

[Implementation](<https://github.com/Microsoft/robust-models-transfer>)

Recorded evaluation configuration: {"source":"Salman","Architecture":"ResNet-50","Iterations":"1000"}. This record does not assert independent reproduction.

### Salman WideResNet-50-2

Parsimon: 42.94. Result ID: 69970b7d5698665ec458ef4b.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68e21e209131cea1ef57.md>)

[Implementation](<https://github.com/Microsoft/robust-models-transfer>)

Recorded evaluation configuration: {"source":"Salman","Architecture":"WideResNet-50-2","Iterations":"1000"}. This record does not assert independent reproduction.

### Debenedetti XCiT-S12

Parsimon: 37.44. Result ID: 69970b7d5698665ec458ef4d.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e741e209131cea3ff08.md>)

Recorded evaluation configuration: {"source":"Debenedetti","Architecture":"XCiT-S12","Iterations":"1000"}. This record does not assert independent reproduction.

### Debenedetti XCiT-M12

Parsimon: 36.04. Result ID: 69970b7d5698665ec458ef4e.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e741e209131cea3ff08.md>)

Recorded evaluation configuration: {"source":"Debenedetti","Architecture":"XCiT-M12","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ViT-S+ConvStem

Parsimon: 35.5. Result ID: 69970b7d5698665ec458ef50.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ViT-S+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Debenedetti XCiT-L12

Parsimon: 35.32. Result ID: 69970b7d5698665ec458ef4f.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e741e209131cea3ff08.md>)

Recorded evaluation configuration: {"source":"Debenedetti","Architecture":"XCiT-L12","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ConvNeXt-T+ConvStem

Parsimon: 34.88. Result ID: 69970b7d5698665ec458ef52.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ConvNeXt-T+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ConvNeXt-S+ConvStem

Parsimon: 33.36. Result ID: 69970b7d5698665ec458ef53.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ConvNeXt-S+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ViT-B+ConvStem

Parsimon: 31.02. Result ID: 69970b7d5698665ec458ef51.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ViT-B+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Liu Swin-B

Parsimon: 30.88. Result ID: 69970b7d5698665ec458ef58.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6eb01e209131cea41754.md>)

Recorded evaluation configuration: {"source":"Liu","Architecture":"Swin-B","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ConvNeXt-B+ConvStem

Parsimon: 30.78. Result ID: 69970b7d5698665ec458ef54.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ConvNeXt-B+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Liu ConvNeXt-B

Parsimon: 30.32. Result ID: 69970b7d5698665ec458ef56.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6eb01e209131cea41754.md>)

Recorded evaluation configuration: {"source":"Liu","Architecture":"ConvNeXt-B","Iterations":"1000"}. This record does not assert independent reproduction.

### Singh ConvNeXt-L+ConvStem

Parsimon: 29.24. Result ID: 69970b7d5698665ec458ef55.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71011e209131cea4e7c0.md>)

[Implementation](<https://github.com/nmndeep/revisiting-at>)

Recorded evaluation configuration: {"source":"Singh","Architecture":"ConvNeXt-L+ConvStem","Iterations":"1000"}. This record does not assert independent reproduction.

### Liu ConvNeXt-L

Parsimon: 28.92. Result ID: 69970b7d5698665ec458ef57.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6eb01e209131cea41754.md>)

Recorded evaluation configuration: {"source":"Liu","Architecture":"ConvNeXt-L","Iterations":"1000"}. This record does not assert independent reproduction.

### Liu Swin-L

Parsimon: 28.18. Result ID: 69970b7d5698665ec458ef59.

[Reporting paper](<https://www.sota2.com/research/paper/692fc3e7d6b0dd334293318c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6eb01e209131cea41754.md>)

Recorded evaluation configuration: {"source":"Liu","Architecture":"Swin-L","Iterations":"1000"}. This record does not assert independent reproduction.

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