Jailbreak Detection on Base64
100AccuracyIMAG
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| IMAGTarget LLM=Llama3-8B2025.12 | 100 | 100 | |
| PPLTarget LLM=Mistral-7B2025.12 | 95 | 95 | |
| IMAGTarget LLM=Mistral-7B2025.12 | 95 | 97 | |
| PPLTarget LLM=Vicuna-7B2025.12 | 95 | 95 | |
| PPLTarget LLM=Vicuna-13B2025.12 | 95 | 95 | |
| PPLTarget LLM=Llama2-7B2025.12 | 95 | 95 | |
| PPLTarget LLM=Llama3-8B2025.12 | 95 | 95 | |
| IMAGTarget LLM=Vicuna-7B2025.12 | 91 | 95 | |
| IMAGTarget LLM=Vicuna-13B2025.12 | 91 | 95 | |
| IMAGTarget LLM=Llama2-7B2025.12 | 60 | 75 | |
| LlamaGTarget LLM=Mistral-7B2025.12 | 50 | 67 | |
| LlamaGTarget LLM=Vicuna-7B2025.12 | 49 | 65 | |
| LlamaGTarget LLM=Llama2-7B2025.12 | 49 | 65 | |
| LlamaGTarget LLM=Vicuna-13B2025.12 | 48 | 64 | |
| LlamaGTarget LLM=Llama3-8B2025.12 | 34 | 51 | |
| Self-ExTarget LLM=Mistral-7B2025.12 | 32 | 49 | |
| Self-ExTarget LLM=Llama2-7B2025.12 | 20 | 32 | |
| Self-ExTarget LLM=Llama3-8B2025.12 | 12 | 21 | |
| Self-ExTarget LLM=Vicuna-7B2025.12 | 1 | 2 | |
| OAPITarget LLM=Mistral-7B2025.12 | 0 | 0 | |
| GradSafeTarget LLM=Mistral-7B2025.12 | 0 | 0 | |
| OAPITarget LLM=Vicuna-7B2025.12 | 0 | 0 | |
| GradSafeTarget LLM=Vicuna-7B2025.12 | 0 | 0 | |
| OAPITarget LLM=Vicuna-13B2025.12 | 0 | 0 | |
| Self-ExTarget LLM=Vicuna-13B2025.12 | 0 | 0 | |
| GradSafeTarget LLM=Vicuna-13B2025.12 | 0 | 0 | |
| OAPITarget LLM=Llama2-7B2025.12 | 0 | 0 | |
| GradSafeTarget LLM=Llama2-7B2025.12 | 0 | 0 | |
| OAPITarget LLM=Llama3-8B2025.12 | 0 | 0 | |
| GradSafeTarget LLM=Llama3-8B2025.12 | 0 | 0 |