# Perturbation-based Jailbreak Attack Defense on Perturbation-based Attack adversarial variants

Source: [Perturbation-based Jailbreak Attack Defense on Perturbation-based Attack adversarial variants](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants>)

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

Data updated: 2026-03-24T17:07:46.877Z

Evaluation of model defense performance against perturbation-based jailbreak attacks, measuring Attack Success Rate (ASR) under various gradient-based adversarial methods (PGD, Auto-PGD, MI-FGSM) at different perturbation budgets (eps=16/255, eps=32/255).

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-based Jailbreak Attack Defense |
| Dataset (including any stated split) | Perturbation-based Attack adversarial variants |
| Metric | ASR (PGD, eps=16/255) |
| 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

- [ASR (PGD, eps=16/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(PGD%2C%20eps%3D16%2F255)>): Higher is better; unit: PERCENT.

- [ASR (PGD, eps=32/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(PGD%2C%20eps%3D32%2F255)>): Higher is better; unit: PERCENT.

- [ASR (Auto-PGD, eps=16/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(Auto-PGD%2C%20eps%3D16%2F255)>): Higher is better; unit: PERCENT.

- [ASR (Auto-PGD, eps=32/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(Auto-PGD%2C%20eps%3D32%2F255)>): Higher is better; unit: PERCENT.

- [ASR (MI-FGSM, eps=16/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(MI-FGSM%2C%20eps%3D16%2F255)>): Higher is better; unit: PERCENT.

- [ASR (MI-FGSM, eps=32/255)](<https://www.sota2.com/research/sota/perturbation-based-jailbreak-attack-defense-on-perturbation-based-attack-adversarial-variants.md?metric=ASR%20(MI-FGSM%2C%20eps%3D32%2F255)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | ASR (PGD, eps=16/255) | Result date | Report paper |
| --- | --- | --- | --- |
| w/o defense | 74.55 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| w/o defense | 70.91 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| w/o defense | 69.09 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| ASTRA | 21.82 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| NullSteer | 19.95 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| ASTRA | 5.45 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| NullSteer | 3.28 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| ASTRA | 1.82 | Not recorded | 69c21ad30636e5bf4afc4d16 |
| NullSteer | 1.74 | Not recorded | 69c21ad30636e5bf4afc4d16 |

### w/o defense

ASR (PGD, eps=16/255): 74.55. Result ID: 69c2c4ca47f34ce924ac813b.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"Qwen2-VL"}. This record does not assert independent reproduction.

### w/o defense

ASR (PGD, eps=16/255): 70.91. Result ID: 69c2c4ca47f34ce924ac8138.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"MiniGPT-4"}. This record does not assert independent reproduction.

### w/o defense

ASR (PGD, eps=16/255): 69.09. Result ID: 69c2c4ca47f34ce924ac813e.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5"}. This record does not assert independent reproduction.

### ASTRA

ASR (PGD, eps=16/255): 21.82. Result ID: 69c2c4ca47f34ce924ac813c.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78541e209131cea77be4.md>)

Recorded evaluation configuration: {"Backbone":"Qwen2-VL"}. This record does not assert independent reproduction.

### NullSteer

ASR (PGD, eps=16/255): 19.95. Result ID: 69c2c4ca47f34ce924ac813d.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"Qwen2-VL"}. This record does not assert independent reproduction.

### ASTRA

ASR (PGD, eps=16/255): 5.45. Result ID: 69c2c4ca47f34ce924ac8139.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78541e209131cea77be4.md>)

Recorded evaluation configuration: {"Backbone":"MiniGPT-4"}. This record does not assert independent reproduction.

### NullSteer

ASR (PGD, eps=16/255): 3.28. Result ID: 69c2c4ca47f34ce924ac813a.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"MiniGPT-4"}. This record does not assert independent reproduction.

### ASTRA

ASR (PGD, eps=16/255): 1.82. Result ID: 69c2c4ca47f34ce924ac813f.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78541e209131cea77be4.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5"}. This record does not assert independent reproduction.

### NullSteer

ASR (PGD, eps=16/255): 1.74. Result ID: 69c2c4ca47f34ce924ac8140.

[Reporting paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c21ad30636e5bf4afc4d16.md>)

Recorded evaluation configuration: {"Backbone":"LLaVA-v1.5"}. This record does not assert independent reproduction.

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