List your company

Defending against gradient-based attacks on Llama3 GCG Attack (test)

9.61ASROurs-Fakecom-t
6.51427.41248.3169.208Nov 1, 2024

Evaluation results

MethodLinks
Ours-Fakecom-tDefense=Fake Completion Defense with Template2024.119.61
Ours-IgnoreDefense=Ignore Defense2024.1112.01
Ours-FakecomDefense=Fake Completion Defense2024.1113.94
SandwichDefense=Sandwich defense baseline2024.1119.23
Ours-EscapeDefense=Escape Defense2024.1119.23
SpotlightDefense=Spotlight defense baseline2024.1119.71
ReminderDefense=Reminder defense baseline2024.1124.51
InstructionalDefense=Instructional defense baseline2024.1128.84
IsolationDefense=Isolation defense baseline2024.1140.38
NoneDefense=No defense baseline2024.1187.01