Privacy Flow Control on CONFAIDE Tier 3 InfoFlow-Control
50LeakageMistral-7B (Drunk language inducement)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Mistral-7B (Drunk language inducement)Variant=RL2026.01 | 50 | 33 | 77 | 80 | |
| Mistral-7B (Drunk language inducement)Variant=FT2026.01 | 57 | 10 | 66 | 78 | |
| Mistral-7BVariant=Base2026.01 | 69 | 10 | 77 | 57 | |
| GPT-3.5Variant=Base2026.01 | 80 | 0 | 65 | 46 | |
| Mistral-7B (Drunk language inducement)Variant=Pr.2026.01 | 81 | 78 | 77 | 67 | |
| GPT-4 (Drunk language inducement)Variant=FT2026.01 | 96 | 75 | 99 | 97 | |
| GPT-4 (Drunk language inducement)Variant=Pr.2026.01 | 97 | 54 | 90 | 80 | |
| LLaMA2-7B (Drunk language inducement)Variant=Pr.2026.01 | 97 | 100 | 100 | 100 | |
| LLaMA2-7B (Drunk language inducement)Variant=RL2026.01 | 97 | 100 | 100 | 100 | |
| GPT-3.5 (Drunk language inducement)Variant=Pr.2026.01 | 98 | 73 | 100 | 100 | |
| LLaMA2-7BVariant=Base2026.01 | 98 | 100 | 100 | 99 | |
| LLaMA2-7B (Drunk language inducement)Variant=FT2026.01 | 98 | 100 | 100 | 100 | |
| LLaMA3-8B (Drunk language inducement)Variant=RL2026.01 | 98 | 100 | 100 | 100 | |
| GPT-4Variant=Base2026.01 | 100 | 6 | 12 | 7 | |
| LLaMA3-8BVariant=Base2026.01 | 100 | 100 | 100 | 100 | |
| LLaMA3-8B (Drunk language inducement)Variant=Pr.2026.01 | 100 | 82 | 100 | 100 | |
| LLaMA3-8B (Drunk language inducement)Variant=FT2026.01 | 100 | 100 | 100 | 100 |