FlipAttack uses four flipping modes to add prompt-derived left-side noise and four variants to guide LLMs to denoise and execute harmful instructions, achieving high attack success rates on black-box models in a single query.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2024 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
FlipAttack: Jailbreak LLMs via Flipping
FlipAttack uses four flipping modes to add prompt-derived left-side noise and four variants to guide LLMs to denoise and execute harmful instructions, achieving high attack success rates on black-box models in a single query.