Jailbreak Defense on Jailbreak Attacks (Specific ASR Metrics)
0GCG ASRSmoothLLM
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SmoothLLMBackbone=LLaMA2-7B-Chat2026.05 | 0 | 0 | 0 | 84 | 24 | |
| RESTABackbone=LLaMA2-7B-Chat2026.05 | 0 | 0 | 0 | 15 | 20 | |
| embedding disruption methodBackbone=LLaMA2-7B-Chat2026.05 | 0 | 0 | 0 | 0 | 10 | |
| embedding disruption methodBackbone=Qwen2.5-7B-Instruct2026.05 | 0 | 4 | 2 | 1 | 29 | |
| embedding disruption methodBackbone=Vicuna-13B2026.05 | 1 | 0 | 11 | 0 | 26 | |
| Perplexity FilterBackbone=LLaMA2-7B-Chat2026.05 | 1 | 0 | 73 | 98 | 31 | |
| Erase-and-CheckBackbone=LLaMA2-7B-Chat2026.05 | 1 | 0 | 25 | 12 | 14 | |
| RESTABackbone=Vicuna-13B2026.05 | 2 | 30 | 44 | 35 | 44 | |
| Perplexity FilterBackbone=Qwen2.5-7B-Instruct2026.05 | 2 | 31 | 71 | 91 | 60 | |
| Perplexity FilterBackbone=Vicuna-13B2026.05 | 3 | 69 | 88 | 92 | 75 | |
| VanillaBackbone=LLaMA2-7B-Chat2026.05 | 3 | 0 | 90 | 98 | 31 | |
| SmoothLLMBackbone=Vicuna-13B2026.05 | 4 | 55 | 68 | 55 | 52 | |
| RESTABackbone=Qwen2.5-7B-Instruct2026.05 | 8 | 22 | 50 | 38 | 49 | |
| Erase-and-CheckBackbone=Qwen2.5-7B-Instruct2026.05 | 10 | 6 | 29 | 24 | 34 | |
| SmoothLLMBackbone=Qwen2.5-7B-Instruct2026.05 | 12 | 25 | 72 | 60 | 53 | |
| Erase-and-CheckBackbone=Vicuna-13B2026.05 | 17 | 6 | 24 | 21 | 37 | |
| VanillaBackbone=Qwen2.5-7B-Instruct2026.05 | 19 | 31 | 85 | 91 | 60 | |
| VanillaBackbone=Vicuna-13B2026.05 | 80 | 69 | 89 | 95 | 75 |