Stealthiness Evaluation on Harmful prompts (evaluated on 3 LLMs and 4 guard LLMs)
3.23Mean PerplexityArtPrompt
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ArtPrompt2024.10 | 3.23 | 1.89 | |
| Ascii2024.10 | 4.13 | 0.49 | |
| base642024.10 | 10.46 | 3.06 | |
| Morse Cipher2024.10 | 11.81 | 2.19 | |
| ReNeLLM2024.10 | 15.56 | 5.69 | |
| Unicode2024.10 | 42.19 | 33.57 | |
| UTF-82024.10 | 42.19 | 33.57 | |
| Origin2024.10 | 49.9 | 51.63 | |
| Caesar Cipher2024.10 | 258.1 | 182.96 | |
| FlipAttack2024.10 | 809.67 | 506.4 |