# Image Classification on CIFAR100 (Adversarial Robustness - Clean & PGD)

Source: [Image Classification on CIFAR100 (Adversarial Robustness - Clean & PGD)](<https://www.sota2.com/research/sota/image-classification-on-cifar100-adversarial-robustness-clean-and-pgd>)

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

Data updated: 2026-04-07T09:42:43.515Z

Evaluation of Image Classification performance on the CIFAR-100 dataset, reporting Clean Accuracy and robustness against PGD adversarial attacks.

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) | CIFAR100 |
| Metric | Clean Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 36 of 36 |
| 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

- [Clean Accuracy](<https://www.sota2.com/research/sota/image-classification-on-cifar100-adversarial-robustness-clean-and-pgd.md?metric=Clean%20Accuracy>): Higher is better; unit: PERCENT.

- [PGD Robust Accuracy](<https://www.sota2.com/research/sota/image-classification-on-cifar100-adversarial-robustness-clean-and-pgd.md?metric=PGD%20Robust%20Accuracy>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Clean Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| RoLI - Full-FT | 76.25 | Not recorded | 68fa72bf1e209131cea585bb |
| RanLI - Full-FT | 74.36 | Not recorded | 68fa72bf1e209131cea585bb |
| TTE | 72.9 | Not recorded | 69799806a2b28e761d87cb52 |
| RoLI - LoRA | 72.87 | Not recorded | 68fa72bf1e209131cea585bb |
| LPF | 72.8 | Not recorded | 69799806a2b28e761d87cb52 |
| RanLI - VPT | 72.37 | Not recorded | 68fa72bf1e209131cea585bb |
| RoLI - Bias | 72.17 | Not recorded | 68fa72bf1e209131cea585bb |
| RanLI - LoRA | 71.98 | Not recorded | 68fa72bf1e209131cea585bb |
| RoLI - Adapter | 71.75 | Not recorded | 68fa72bf1e209131cea585bb |
| RanLI - Adapter | 71.57 | Not recorded | 68fa72bf1e209131cea585bb |
| RanLI - Bias | 71.5 | Not recorded | 68fa72bf1e209131cea585bb |
| TTC | 71 | Not recorded | 69799806a2b28e761d87cb52 |
| TWINS-AT | 70.72 | Not recorded | 68fa72bf1e209131cea585bb |
| HD | 67.9 | Not recorded | 69799806a2b28e761d87cb52 |
| RanLI - Linear | 67.65 | Not recorded | 68fa72bf1e209131cea585bb |
| AutoLoRa | 66.2 | Not recorded | 68fa72bf1e209131cea585bb |
| R-TPT | 65.7 | Not recorded | 69799806a2b28e761d87cb52 |
| CLIP | 65 | Not recorded | 69799806a2b28e761d87cb52 |
| HPT-GPD | 64.83 | Not recorded | 69705d96e25ad2fa9d1ed1a8 |
| Anti-Adv | 64.7 | Not recorded | 69799806a2b28e761d87cb52 |
| CSR | 62.9 | Not recorded | 69799806a2b28e761d87cb52 |
| CLIP | 60.39 | Not recorded | 69705d96e25ad2fa9d1ed1a8 |
| CLIP-FT | 59.51 | Not recorded | 6923e676d6b0dd334292eee0 |
| FT-Standard | 58.78 | Not recorded | 69705d96e25ad2fa9d1ed1a8 |
| TTE | 58.61 | Not recorded | 6923e676d6b0dd334292eee0 |
| CLIP | 57.14 | Not recorded | 6923e676d6b0dd334292eee0 |
| TTC | 56.34 | Not recorded | 6923e676d6b0dd334292eee0 |
| R-TPT | 52.69 | Not recorded | 6923e676d6b0dd334292eee0 |
| ATAC | 52.65 | Not recorded | 6923e676d6b0dd334292eee0 |
| PMG | 52.28 | Not recorded | 69705d96e25ad2fa9d1ed1a8 |
| FARE | 47.38 | Not recorded | 6923e676d6b0dd334292eee0 |
| FARE | 45.9 | Not recorded | 69799806a2b28e761d87cb52 |
| TeCoA | 44 | Not recorded | 69799806a2b28e761d87cb52 |
| PMG-AFT | 41.51 | Not recorded | 6923e676d6b0dd334292eee0 |
| FT-TeCoA | 41.01 | Not recorded | 69705d96e25ad2fa9d1ed1a8 |
| TeCoA | 36.3 | Not recorded | 6923e676d6b0dd334292eee0 |

### RoLI - Full-FT

Clean Accuracy: 76.25. Result ID: 6992775a3a6c3148490f2799.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Implementation](<https://github.com/DongXzz/RoLI>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Robust Linear Initialization","Finetuning Protocol":"Full Finetuning"}. This record does not assert independent reproduction.

### RanLI - Full-FT

Clean Accuracy: 74.36. Result ID: 6992775a3a6c3148490f2793.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa671d1e209131cea145d5.md>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"Full Finetuning"}. This record does not assert independent reproduction.

### TTE

Clean Accuracy: 72.9. Result ID: 69a17d1735dfcabb00e76291.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a711e209131cea2817e.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### RoLI - LoRA

Clean Accuracy: 72.87. Result ID: 6992775a3a6c3148490f279b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Implementation](<https://github.com/DongXzz/RoLI>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Robust Linear Initialization","Finetuning Protocol":"LoRA"}. This record does not assert independent reproduction.

### LPF

Clean Accuracy: 72.8. Result ID: 69a17d1735dfcabb00e7628e.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### RanLI - VPT

Clean Accuracy: 72.37. Result ID: 6992775a3a6c3148490f2797.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d181e209131cea37ae9.md>)

[Implementation](<https://www.sota2.com/github.com/kmnp/vpt>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"Visual Prompt Tuning"}. This record does not assert independent reproduction.

### RoLI - Bias

Clean Accuracy: 72.17. Result ID: 6992775a3a6c3148490f279c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Implementation](<https://github.com/DongXzz/RoLI>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Robust Linear Initialization","Finetuning Protocol":"Bias-only"}. This record does not assert independent reproduction.

### RanLI - LoRA

Clean Accuracy: 71.98. Result ID: 6992775a3a6c3148490f2795.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"LoRA"}. This record does not assert independent reproduction.

### RoLI - Adapter

Clean Accuracy: 71.75. Result ID: 6992775a3a6c3148490f279a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Implementation](<https://github.com/DongXzz/RoLI>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Robust Linear Initialization","Finetuning Protocol":"Adapter"}. This record does not assert independent reproduction.

### RanLI - Adapter

Clean Accuracy: 71.57. Result ID: 6992775a3a6c3148490f2794.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa65e31e209131cea0d324.md>)

[Implementation](<https://github.com/google-research/adapter-bert>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"Adapter"}. This record does not assert independent reproduction.

### RanLI - Bias

Clean Accuracy: 71.5. Result ID: 6992775a3a6c3148490f2796.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"Bias-only"}. This record does not assert independent reproduction.

### TTC

Clean Accuracy: 71. Result ID: 69a17d1735dfcabb00e76292.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa774a1e209131cea724cc.md>)

[Implementation](<https://www.sota2.com/here>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### TWINS-AT

Clean Accuracy: 70.72. Result ID: 6992775a3a6c3148490f2791.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6ed11e209131cea4265b.md>)

[Implementation](<https://github.com/ziquanliu/CVPR2023-TWINS>)

Recorded evaluation configuration: {"Backbone":"ResNet-50"}. This record does not assert independent reproduction.

### HD

Clean Accuracy: 67.9. Result ID: 69a17d1735dfcabb00e7628f.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a1d1e209131cea26045.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### RanLI - Linear

Clean Accuracy: 67.65. Result ID: 6992775a3a6c3148490f2798.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa671d1e209131cea145d5.md>)

Recorded evaluation configuration: {"Backbone":"Swin Transformer","Initialization":"Random Initialization","Finetuning Protocol":"Linear Probing"}. This record does not assert independent reproduction.

### AutoLoRa

Clean Accuracy: 66.2. Result ID: 6992775a3a6c3148490f2792.

[Reporting paper](<https://www.sota2.com/research/paper/68fa72bf1e209131cea585bb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70b91e209131cea4cd55.md>)

Recorded evaluation configuration: {"Backbone":"ResNet-50"}. This record does not assert independent reproduction.

### R-TPT

Clean Accuracy: 65.7. Result ID: 69a17d1735dfcabb00e7628d.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7add1e209131cea8590d.md>)

[Implementation](<https://github.com/TomSheng21/R-TPT>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### CLIP

Clean Accuracy: 65. Result ID: 69a17d1735dfcabb00e7628a.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69581e209131cea21ad0.md>)

[Implementation](<https://github.com/OpenAI/CLIP>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Original"}. This record does not assert independent reproduction.

### HPT-GPD

Clean Accuracy: 64.83. Result ID: 69a17dcc35dfcabb00e76582.

[Reporting paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

Recorded evaluation configuration: {"Backbone":"ViT-B/32","Proxy CLIP":"ViT-B/16","Attack Protocol":"PGD-10","Training time per epoch (s)":"1486"}. This record does not assert independent reproduction.

### Anti-Adv

Clean Accuracy: 64.7. Result ID: 69a17d1735dfcabb00e76290.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b7f1e209131cea2e9af.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### CSR

Clean Accuracy: 62.9. Result ID: 69a17d1735dfcabb00e76293.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Test-Time Defense"}. This record does not assert independent reproduction.

### CLIP

Clean Accuracy: 60.39. Result ID: 69a17dcc35dfcabb00e7657e.

[Reporting paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69581e209131cea21ad0.md>)

[Implementation](<https://github.com/OpenAI/CLIP>)

Recorded evaluation configuration: {"Backbone":"ViT-B/32","Proxy CLIP":"ViT-B/16","Attack Protocol":"PGD-10","Training time per epoch (s)":"0"}. This record does not assert independent reproduction.

### CLIP-FT

Clean Accuracy: 59.51. Result ID: 69d4d19047f34ce924bcfcfb.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

Recorded evaluation configuration: {"Defense Category":"Adversarial Finetuning","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### FT-Standard

Clean Accuracy: 58.78. Result ID: 69a17dcc35dfcabb00e7657f.

[Reporting paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

Recorded evaluation configuration: {"Backbone":"ViT-B/32","Proxy CLIP":"ViT-B/16","Attack Protocol":"PGD-10","Training time per epoch (s)":"406"}. This record does not assert independent reproduction.

### TTE

Clean Accuracy: 58.61. Result ID: 69d4d19047f34ce924bcfcff.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6a711e209131cea2817e.md>)

Recorded evaluation configuration: {"Defense Category":"Test-time Defense","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### CLIP

Clean Accuracy: 57.14. Result ID: 69d4d19047f34ce924bcfcfa.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69581e209131cea21ad0.md>)

[Implementation](<https://github.com/OpenAI/CLIP>)

Recorded evaluation configuration: {"Defense Category":"None","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### TTC

Clean Accuracy: 56.34. Result ID: 69d4d19047f34ce924bcfd00.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa774a1e209131cea724cc.md>)

[Implementation](<https://www.sota2.com/here>)

Recorded evaluation configuration: {"Defense Category":"Test-time Defense","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### R-TPT

Clean Accuracy: 52.69. Result ID: 69d4d19047f34ce924bcfd01.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7add1e209131cea8590d.md>)

[Implementation](<https://github.com/TomSheng21/R-TPT>)

Recorded evaluation configuration: {"Defense Category":"Test-time Defense","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### ATAC

Clean Accuracy: 52.65. Result ID: 69d4d19047f34ce924bcfd02.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

Recorded evaluation configuration: {"Defense Category":"Test-time Defense","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### PMG

Clean Accuracy: 52.28. Result ID: 69a17dcc35dfcabb00e76581.

[Reporting paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72df1e209131cea5912e.md>)

Recorded evaluation configuration: {"Backbone":"ViT-B/32","Proxy CLIP":"ViT-B/16","Attack Protocol":"PGD-10","Training time per epoch (s)":"1817"}. This record does not assert independent reproduction.

### FARE

Clean Accuracy: 47.38. Result ID: 69d4d19047f34ce924bcfcfe.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa738c1e209131cea5d24a.md>)

Recorded evaluation configuration: {"Defense Category":"Adversarial Finetuning","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### FARE

Clean Accuracy: 45.9. Result ID: 69a17d1735dfcabb00e7628c.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa738c1e209131cea5d24a.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Adversarial Fine-tuning"}. This record does not assert independent reproduction.

### TeCoA

Clean Accuracy: 44. Result ID: 69a17d1735dfcabb00e7628b.

[Reporting paper](<https://www.sota2.com/research/paper/69799806a2b28e761d87cb52.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f271e209131cea44177.md>)

Recorded evaluation configuration: {"Backbone":"CLIP-L/14","Method category":"Adversarial Fine-tuning"}. This record does not assert independent reproduction.

### PMG-AFT

Clean Accuracy: 41.51. Result ID: 69d4d19047f34ce924bcfcfd.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72df1e209131cea5912e.md>)

Recorded evaluation configuration: {"Defense Category":"Adversarial Finetuning","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

### FT-TeCoA

Clean Accuracy: 41.01. Result ID: 69a17dcc35dfcabb00e76580.

[Reporting paper](<https://www.sota2.com/research/paper/69705d96e25ad2fa9d1ed1a8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f271e209131cea44177.md>)

Recorded evaluation configuration: {"Backbone":"ViT-B/32","Proxy CLIP":"ViT-B/16","Attack Protocol":"PGD-10","Training time per epoch (s)":"1112"}. This record does not assert independent reproduction.

### TeCoA

Clean Accuracy: 36.3. Result ID: 69d4d19047f34ce924bcfcfc.

[Reporting paper](<https://www.sota2.com/research/paper/6923e676d6b0dd334292eee0.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f271e209131cea44177.md>)

Recorded evaluation configuration: {"Defense Category":"Adversarial Finetuning","Attack Type":"10-step PGD","Epsilon":"4/255"}. This record does not assert independent reproduction.

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