Input Collision+ Attack on Input Reconstruction Dataset
1BERTScore (First)Original
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalModel=LLaMA-3.2-3B-Instruct2025.08 | 1 | 1 | 1 | 0.994 | 0.994 | 0.994 | |
| OriginalModel=LLaMA-3.1-8B2025.08 | 1 | 1 | 1 | 0.994 | 0.994 | 0.994 | |
| OriginalModel=Qwen2.5-Math-7B2025.08 | 1 | 0.983 | 0.996 | 1 | 0.977 | 0.996 | |
| (10^8, 10^{-5})-DPModel=LLaMA-3.2-3B-Instruct2025.08 | 0.967 | 0.938 | 1 | 0.951 | 0.907 | 0.994 | |
| (10^8, 10^{-5})-DPModel=LLaMA-3.1-8B2025.08 | 0.639 | 0.986 | 0.999 | 0.639 | 0.947 | 0.994 | |
| (10^8, 10^{-5})-DPModel=Qwen2.5-Math-7B2025.08 | 0.331 | 0.432 | 0.373 | 0.336 | 0.445 | 0.404 | |
| KV-CloakModel=Qwen2.5-Math-7B2025.08 | 0.112 | 0.112 | 0.113 | 0 | 0 | 0 | |
| KV-CloakModel=LLaMA-3.1-8B2025.08 | 0.041 | 0.041 | 0.041 | 0.003 | 0.003 | 0.003 | |
| KV-CloakModel=LLaMA-3.2-3B-Instruct2025.08 | 0.033 | 0.033 | 0.033 | 0.042 | 0.042 | 0.042 |