Injection Attack on Input Reconstruction (All)
76.5BERTScoreOriginal
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OriginalModel=LLaMA-7B2025.08 | 76.5 | 68.7 | |
| OriginalModel=LLaMA-7B, Protect Type=Original2025.08 | 76.5 | 68.7 | |
| OriginalModel=LLaMA-3.1-8B2025.08 | 61.6 | 44.7 | |
| OriginalModel=LLaMA-3.1-8B-Distilled2025.08 | 61 | 42.1 | |
| OriginalModel=LLaMA-3.1-8B-Distilled, Protect Type=Original2025.08 | 61 | 42.1 | |
| OriginalModel=LLaMA-3.2-1B2025.08 | 54.4 | 31.5 | |
| OriginalModel=LLaMA-3.2-1B, Protect Type=Original2025.08 | 54.4 | 31.5 | |
| OriginalModel=LLaMA-3.2-3B-Instruct2025.08 | 54 | 32.4 | |
| OriginalModel=Qwen2.5-Math-7B2025.08 | 42.2 | 28.6 | |
| (10^8, 10^-5)-DPModel=LLaMA-3.2-1B2025.08 | 39.3 | 24.8 | |
| (10^8, 10^-5)-DPModel=LLaMA-3.2-1B, Protect Type=(10^8, 10^-5)-DP2025.08 | 39.3 | 24.8 | |
| (10^8, 10^{-5})-DPModel=Qwen2.5-Math-7B2025.08 | 32.5 | 20.8 | |
| (10^8, 10^{-5})-DPModel=LLaMA-3.2-3B-Instruct2025.08 | 12.9 | 3.2 | |
| (10^8, 10^-5)-DPModel=LLaMA-3.1-8B-Distilled2025.08 | 11.8 | 4.9 | |
| (10^8, 10^-5)-DPModel=LLaMA-3.1-8B-Distilled, Protect Type=(10^8, 10^-5)-DP2025.08 | 11.8 | 4.9 | |
| (10^8, 10^{-5})-DPModel=LLaMA-3.1-8B2025.08 | 11.5 | 5.7 | |
| KV-CloakModel=LLaMA-3.1-8B-Distilled2025.08 | 8.8 | 0 | |
| KV-CloakModel=LLaMA-3.2-3B-Instruct2025.08 | 8.8 | 0 | |
| KV-CloakModel=LLaMA-3.1-8B-Distilled, Protect Type=KV-Cloak2025.08 | 8.8 | 0 | |
| (10^8, 10^-5)-DPModel=LLaMA-7B2025.08 | 8.5 | 0.9 | |
| (10^8, 10^-5)-DPModel=LLaMA-7B, Protect Type=(10^8, 10^-5)-DP2025.08 | 8.5 | 0.9 | |
| KV-CloakModel=LLaMA-3.1-8B2025.08 | 8.4 | 0 | |
| KV-CloakModel=LLaMA-7B2025.08 | 8.2 | 0 | |
| KV-CloakModel=LLaMA-7B, Protect Type=KV-Cloak2025.08 | 8.2 | 0 | |
| KV-CloakModel=LLaMA-3.2-1B2025.08 | 7.9 | 0 | |
| KV-CloakModel=LLaMA-3.2-1B, Protect Type=KV-Cloak2025.08 | 7.9 | 0 | |
| KV-CloakModel=Qwen2.5-Math-7B2025.08 | 7.5 | 0 |