Jailbreak Detection on Average of six attacks
0Avg Success RateGradSafe
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GradSafeTarget LLM=Vicuna-7B2025.12 | 0 | — | |
| Self-ExTarget LLM=Vicuna-13B2025.12 | 0 | — | |
| GradSafeTarget LLM=Vicuna-13B2025.12 | 0 | — | |
| Self-ExTarget LLM=Vicuna-7B2025.12 | 2 | — | |
| OAPITarget LLM=Vicuna-7B2025.12 | 5 | — | |
| OAPITarget LLM=Vicuna-13B2025.12 | 5 | — | |
| OAPITarget LLM=Llama2-7B2025.12 | 5 | — | |
| OAPITarget LLM=Mistral-7B2025.12 | 6 | — | |
| OAPITarget LLM=Llama3-8B2025.12 | 6 | — | |
| GradSafeTarget LLM=Llama3-8B2025.12 | 7 | — | |
| SelfGraderQ=10, Latency (Sec.)=0.74, Memory Overhead (MB)=2765.652026.04 | 7.45 | 18.5 | |
| SelfDefendDefense Protocol=Intent, Latency (Sec.)=2.14, Memory Overhead (MB)=13194.112026.04 | 8.95 | 29.96 | |
| GradSafeTarget LLM=Mistral-7B2025.12 | 12 | — | |
| SelfDefendDefense Protocol=Direct, Latency (Sec.)=0.84, Memory Overhead (MB)=13183.912026.04 | 13.68 | 36.41 | |
| Self-ExTarget LLM=Llama3-8B2025.12 | 19 | — | |
| PPLTarget LLM=Mistral-7B2025.12 | 22 | — | |
| Prompt GuardLatency (Sec.)=15.11, Memory Overhead (MB)=1064.732026.04 | 23.51 | 46.53 | |
| PPLTarget LLM=Vicuna-7B2025.12 | 27 | — | |
| Self-ExTarget LLM=Llama2-7B2025.12 | 28 | — | |
| PPLTarget LLM=Vicuna-13B2025.12 | 30 | — | |
| PPLTarget LLM=Llama2-7B2025.12 | 31 | — | |
| Llama GuardDeployment Stage=Post, Latency (Sec.)=0.56, Memory Overhead (MB)=13753.872026.04 | 35.77 | 77.73 | |
| Llama GuardDeployment Stage=Pre, Latency (Sec.)=0.49, Memory Overhead (MB)=13649.822026.04 | 36.05 | 74.86 | |
| Perplexity FilterLatency (Sec.)=0.17, Memory Overhead (MB)=13214.012026.04 | 40.61 | 84 | |
| PPLTarget LLM=Llama3-8B2025.12 | 50 | — | |
| LlamaGTarget LLM=Llama2-7B2025.12 | 52 | — | |
| Self-ExTarget LLM=Mistral-7B2025.12 | 54 | — | |
| Vicuna-13B-v1.5Defense=No Defense, Latency (Sec.)=3.202026.04 | 54.94 | — | |
| LlamaGTarget LLM=Llama3-8B2025.12 | 59 | — | |
| GradSafeTarget LLM=Llama2-7B2025.12 | 64 | — | |
| LlamaGTarget LLM=Vicuna-13B2025.12 | 73 | — | |
| LlamaGTarget LLM=Vicuna-7B2025.12 | 74 | — | |
| IMAGTarget LLM=Llama2-7B2025.12 | 74 | — | |
| LlamaGTarget LLM=Mistral-7B2025.12 | 75 | — | |
| IMAGTarget LLM=Vicuna-7B2025.12 | 81 | — | |
| IMAGTarget LLM=Llama3-8B2025.12 | 85 | — | |
| IMAGTarget LLM=Vicuna-13B2025.12 | 93 | — | |
| IMAGTarget LLM=Mistral-7B2025.12 | 94 | — |