# Perturbation Test (Impact on Accuracy) on CIFAR-10

Source: [Perturbation Test (Impact on Accuracy) on CIFAR-10](<https://www.sota2.com/research/sota/perturbation-test-impact-on-accuracy-on-cifar-10>)

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

Data updated: 2026-02-26T05:34:24.027Z

Evaluation of model sensitivity to specific input perturbations on the CIFAR-10 dataset, measuring the resulting shift in predicted labels relative to ground truth labels.

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 | Perturbation Test (Impact on Accuracy) |
| Dataset (including any stated split) | CIFAR-10 |
| Metric | Accuracy (Predicted Neg) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 9 of 9 |
| 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 (Predicted Neg)](<https://www.sota2.com/research/sota/perturbation-test-impact-on-accuracy-on-cifar-10.md?metric=Accuracy%20(Predicted%20Neg)>): Higher is better; unit: PERCENT.

- [Accuracy (Predicted Pos)](<https://www.sota2.com/research/sota/perturbation-test-impact-on-accuracy-on-cifar-10.md?metric=Accuracy%20(Predicted%20Pos)>): Lower is better; unit: PERCENT.

- [Accuracy (GT Neg)](<https://www.sota2.com/research/sota/perturbation-test-impact-on-accuracy-on-cifar-10.md?metric=Accuracy%20(GT%20Neg)>): Higher is better; unit: PERCENT.

- [Accuracy (GT Pos)](<https://www.sota2.com/research/sota/perturbation-test-impact-on-accuracy-on-cifar-10.md?metric=Accuracy%20(GT%20Pos)>): Lower is better; unit: PERCENT.

## Recorded results

| Method | Accuracy (Predicted Neg) | Result date | Report paper |
| --- | --- | --- | --- |
| TokenTM | 74.9 | Not recorded | 68fa729e1e209131cea57935 |
| LRP | 72.62 | Not recorded | 68fa729e1e209131cea57935 |
| GAE | 71.67 | Not recorded | 68fa729e1e209131cea57935 |
| Trans. Attr. | 67.65 | Not recorded | 68fa729e1e209131cea57935 |
| Rollout | 64.88 | Not recorded | 68fa729e1e209131cea57935 |
| Grad-CAM | 61.24 | Not recorded | 68fa729e1e209131cea57935 |
| Raw Att. | 58.46 | Not recorded | 68fa729e1e209131cea57935 |
| Cons. LRP | 44.92 | Not recorded | 68fa729e1e209131cea57935 |
| ATTCAT | 41.78 | Not recorded | 68fa729e1e209131cea57935 |

### TokenTM

Accuracy (Predicted Neg): 74.9. Result ID: 6992643c3a6c3148490ea277.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### LRP

Accuracy (Predicted Neg): 72.62. Result ID: 6992643c3a6c3148490ea270.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa625f1e209131ce9f8e34.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### GAE

Accuracy (Predicted Neg): 71.67. Result ID: 6992643c3a6c3148490ea276.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69a11e209131cea233a8.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### Trans. Attr.

Accuracy (Predicted Neg): 67.65. Result ID: 6992643c3a6c3148490ea272.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69b41e209131cea2394f.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### Rollout

Accuracy (Predicted Neg): 64.88. Result ID: 6992643c3a6c3148490ea274.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa67ae1e209131cea17be6.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### Grad-CAM

Accuracy (Predicted Neg): 61.24. Result ID: 6992643c3a6c3148490ea26f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66cb1e209131cea12568.md>)

[Implementation](<https://github.com/ramprs/grad-cam/>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### Raw Att.

Accuracy (Predicted Neg): 58.46. Result ID: 6992643c3a6c3148490ea273.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa665a1e209131cea0fbee.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### Cons. LRP

Accuracy (Predicted Neg): 44.92. Result ID: 6992643c3a6c3148490ea271.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6cd51e209131cea3613b.md>)

[Implementation](<https://github.com/AmeenAli/XAI_Transformers>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

### ATTCAT

Accuracy (Predicted Neg): 41.78. Result ID: 6992643c3a6c3148490ea275.

[Reporting paper](<https://www.sota2.com/research/paper/68fa729e1e209131cea57935.md>)

Recorded evaluation configuration: {"Backbone":"DeiT"}. This record does not assert independent reproduction.

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