Sign In

FlipAttack: Jailbreak LLMs via Flipping

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Yue Liu, Xiaoxin He, Miao Xiong, Jinlan Fu, Shumin Deng, Yingwei Ma, Jiaheng Zhang, Bryan Hooi

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ์ž๋™ํšŒ๊ท€์  ํŠน์„ฑ์„ ํ™œ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ์˜ ์™ผ์ชฝ ๋ถ€๋ถ„์— ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ ์œ ํ•ดํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํƒ์ง€ํ•˜๊ธฐ ์–ด๋ ต๊ฒŒ ๋งŒ๋“œ๋Š” FlipAttack์ด๋ผ๋Š” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ํƒˆ์˜ฅ ๊ณต๊ฒฉ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ง„์€ LLM์ด ํ…์ŠคํŠธ๋ฅผ ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์ดํ•ดํ•˜๋ฉฐ, ์™ผ์ชฝ ๋…ธ์ด์ฆˆ์— ์ทจ์•ฝํ•˜๋‹ค๋Š” ์ ์„ ๋ฐœ๊ฒฌํ•˜๊ณ , ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ 4๊ฐ€์ง€ ๋ณ€ํ˜• ๋ชจ๋“œ๋ฅผ ๊ฐœ๋ฐœํ•˜์—ฌ LLM์ด ์œ ํ•ดํ•œ ํ–‰๋™์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณต๊ฒฉ์€ ๋‹จ ํ•˜๋‚˜์˜ ์ฟผ๋ฆฌ๋งŒ์œผ๋กœ๋„ ๋ธ”๋ž™๋ฐ•์Šค LLM์„ ํƒˆ์˜ฅ์‹œํ‚ฌ ์ˆ˜ ์žˆ์œผ๋ฉฐ, GPT-4o์™€ ๊ฐ™์€ ์ตœ์‹  ๋ชจ๋ธ์—์„œ๋„ ๋†’์€ ์„ฑ๊ณต๋ฅ ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
LLM์˜ ์ˆœ์ฐจ์  ์ฒ˜๋ฆฌ ๋ฐฉ์‹์˜ ์ทจ์•ฝ์ : FlipAttack์€ LLM์ด ํ…์ŠคํŠธ๋ฅผ ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐฉ์‹์— ์ดˆ์ ์„ ๋งž์ถฐ, ํ…์ŠคํŠธ ์•ž๋ถ€๋ถ„์— ๋ฏธ๋ฌ˜ํ•œ ๋…ธ์ด์ฆˆ๋ฅผ ์‚ฝ์ž…ํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋„ ๋ชจ๋ธ์˜ ๋ฐฉ์–ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์šฐํšŒํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
ํšจ์œจ์ ์ด๊ณ  ๋ณดํŽธ์ ์ธ ๊ณต๊ฒฉ ๋ฐฉ๋ฒ•: ๋‹จ ํ•œ ๋ฒˆ์˜ ์ฟผ๋ฆฌ๋งŒ์œผ๋กœ ๋†’์€ ์„ฑ๊ณต๋ฅ ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์—ฌ๋Ÿฌ LLM ๋ชจ๋ธ๊ณผ ๋‹ค์–‘ํ•œ ๊ฐ€๋“œ๋ ˆ์ผ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํšจ๊ณผ์ ์ž„์ด ์ž…์ฆ๋˜์–ด LLM ๋ณด์•ˆ์— ๋Œ€ํ•œ ์‹ฌ๊ฐํ•œ ์šฐ๋ ค๋ฅผ ์ œ๊ธฐํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋ฐฉ์–ด ๊ธฐ๋ฒ• ๊ฐœ๋ฐœ์˜ ํ•„์š”์„ฑ: ํ˜„์žฌ LLM์˜ ๊ฐ€๋“œ๋ ˆ์ผ ์‹œ์Šคํ…œ์ด ํ…์ŠคํŠธ ์•ž๋ถ€๋ถ„์˜ ๋…ธ์ด์ฆˆ์— ์ทจ์•ฝํ•˜๋‹ค๋Š” ์ ์ด ๋“œ๋Ÿฌ๋‚ฌ์œผ๋ฉฐ, ์ด์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ์–ด ๊ธฐ๋ฒ• ๊ฐœ๋ฐœ์ด ์‹œ๊ธ‰ํžˆ ์š”๊ตฌ๋ฉ๋‹ˆ๋‹ค.
๐Ÿ‘
Made with Slashpage