Defending against Gradient-based Attacks on Llama3 AutoDAN Attack (test)
10.57ASROurs-Fakecom-t
Evaluation Results
| Method | Links | |
|---|---|---|
| Ours-Fakecom-tDefense=Fake Completion Defense with Template2024.11 | 10.57 | |
| Ours-FakecomDefense=Fake Completion Defense2024.11 | 14.9 | |
| Ours-IgnoreDefense=Ignore Defense2024.11 | 16.34 | |
| SpotlightDefense=Spotlight defense baseline2024.11 | 24.51 | |
| Ours-EscapeDefense=Escape Defense2024.11 | 38.94 | |
| SandwichDefense=Sandwich defense baseline2024.11 | 39.42 | |
| ReminderDefense=Reminder defense baseline2024.11 | 51.44 | |
| InstructionalDefense=Instructional defense baseline2024.11 | 52.88 | |
| IsolationDefense=Isolation defense baseline2024.11 | 54.32 | |
| NoneDefense=No defense baseline2024.11 | 68.75 |