Privacy-preserving payload sanitization on 100 medical-booking workflows (test)
97LRPEP
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PEPLocal Backbone=Llama 3.2 (3B), Downstream LLM=GPT-4o-mini2026.05 | 97 | 63.2 | 99 | 6.5 | 3.1 | 6.5 | 39.6 | 49.9 | 86 | 14.3 | 2.028 | |
| PEPLocal Backbone=Llama 3.2 (3B), Downstream LLM=Claude Haiku 4.52026.05 | 97 | 63.2 | 99 | 6.5 | 3.1 | 6.5 | 39.6 | 16.9 | 6 | 14.3 | 2.028 | |
| PEPLocal Backbone=Llama 3.2 (3B), Downstream LLM=Gemini 2.5 Flash2026.05 | 97 | 63.2 | 99 | 6.5 | 3.1 | 6.5 | 39.6 | 58.8 | 85 | 14.3 | 2.028 | |
| VanillaLocal Backbone=Llama 3.2 (3B), Downstream LLM=GPT-4o-mini2026.05 | 96 | 81.2 | 100 | 9.5 | 5.1 | 9.9 | 64.3 | — | 87 | — | — | |
| VanillaLocal Backbone=Llama 3.2 (3B), Downstream LLM=Claude Haiku 4.52026.05 | 96 | 81.2 | 100 | 9.5 | 5.1 | 9.9 | 64.3 | — | 9 | — | — | |
| VanillaLocal Backbone=Llama 3.2 (3B), Downstream LLM=Gemini 2.5 Flash2026.05 | 96 | 81.2 | 100 | 9.5 | 5.1 | 9.9 | 64.3 | — | 89 | — | — | |
| PresidioLocal Backbone=Llama 3.2 (3B), Downstream LLM=GPT-4o-mini2026.05 | 94 | 58.2 | 100 | 4.9 | 1.5 | 5.1 | 29.3 | 25.5 | 73 | 22.1 | 0.042 | |
| PresidioLocal Backbone=Llama 3.2 (3B), Downstream LLM=Claude Haiku 4.52026.05 | 94 | 58.2 | 100 | 4.9 | 1.5 | 5.1 | 29.3 | 21.3 | 1 | 22.1 | 0.042 | |
| PresidioLocal Backbone=Llama 3.2 (3B), Downstream LLM=Gemini 2.5 Flash2026.05 | 94 | 58.2 | 100 | 4.9 | 1.5 | 5.1 | 29.3 | 7.5 | 63 | 22.1 | 0.042 | |
| PrivScopeLocal Backbone=Llama 3.2 (3B), Downstream LLM=GPT-4o-mini2026.05 | 88 | 56 | 92 | 3.3 | 0.3 | 3.1 | 23.1 | 54.1 | 82 | 17.8 | 3.13 | |
| PrivScopeLocal Backbone=Llama 3.2 (3B), Downstream LLM=Claude Haiku 4.52026.05 | 88 | 56 | 92 | 3.3 | 0.3 | 3.1 | 23.1 | 24.7 | 17 | 17.8 | 3.13 | |
| PrivScopeLocal Backbone=Llama 3.2 (3B), Downstream LLM=Gemini 2.5 Flash2026.05 | 88 | 56 | 92 | 3.3 | 0.3 | 3.1 | 23.1 | 69.6 | 84 | 17.8 | 3.13 |