Internal memory extraction attack detection on EHRAgent
1AUROCMemPot
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MemPotAttack=IKEA2026.02 | 1 | 0.98 | 100 | 0 | |
| MemPotAttack=MEXTRA_Cosine2026.02 | 0.99 | 0.97 | 100 | 0 | |
| MemPotAttack=MEXTRA_Edit2026.02 | 0.97 | 0.86 | 99 | 0 | |
| MemPotAttack=MEXTRA_General2026.02 | 0.94 | 0.81 | 96 | 0 | |
| AgentAttack=MEXTRA_Edit2026.02 | 0.67 | 0 | 14 | 0.88 | |
| ControlNetAttack=MEXTRA_General2026.02 | 0.63 | 0.02 | 8 | 0.08 | |
| ControlNetAttack=MEXTRA_Cosine2026.02 | 0.59 | 0 | 14 | 0.08 | |
| AgentAttack=MEXTRA_General2026.02 | 0.58 | 0 | 6 | 0.91 | |
| AgentAttack=MEXTRA_Cosine2026.02 | 0.56 | 0 | 4 | 0.96 | |
| ControlNetAttack=MEXTRA_Edit2026.02 | 0.53 | 0.02 | 17 | 0.08 | |
| AgentAttack=IKEA2026.02 | 0.51 | 0 | 2 | 1.01 | |
| ControlNetAttack=IKEA2026.02 | 0.4 | 0 | 0 | 0.1 |