Sign In

Jailbreaking on Text-to-Video Models via Scene Splitting Strategy

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

์ €์ž

Wonjun Lee, Haon Park, Doehyeon Lee, Bumsub Ham, Suhyun Kim

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ์—ฐ๊ตฌ๋Š” ํ…์ŠคํŠธ-๋น„๋””์˜ค(T2V) ๋ชจ๋ธ์˜ ์•ˆ์ „ ์ทจ์•ฝ์ ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด SceneSplit์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๋ธ”๋ž™๋ฐ•์Šค ํƒˆ์˜ฅ(jailbreak) ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. SceneSplit์€ ์œ ํ•ดํ•œ ์„œ์‚ฌ๋ฅผ ๊ฐœ๋ณ„์ ์œผ๋กœ๋Š” ๋ฌดํ•ดํ•œ ์—ฌ๋Ÿฌ ์žฅ๋ฉด์œผ๋กœ ๋ถ„ํ• ํ•˜๊ณ , ์ด ์žฅ๋ฉด๋“ค์˜ ์ˆœ์ฐจ์  ๊ฒฐํ•ฉ์„ ํ†ตํ•ด ์•ˆ์ „ํ•œ ์ƒ์„ฑ ๊ณต๊ฐ„์„ ์œ ํ•ดํ•œ ์˜์—ญ์œผ๋กœ ์œ ๋„ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ณต์ ์ธ ์žฅ๋ฉด ์กฐ์ž‘๊ณผ ์„ฑ๊ณต์ ์ธ ๊ณต๊ฒฉ ํŒจํ„ด ์žฌ์‚ฌ์šฉ ์ „๋žต์„ ํ†ตํ•ด T2V ๋ชจ๋ธ์˜ ์•ˆ์ „ ํ•„ํ„ฐ๋ฅผ ์šฐํšŒํ•˜์—ฌ ๋†’์€ ์„ฑ๊ณต๋ฅ ๋กœ ์œ ํ•ดํ•œ ๋น„๋””์˜ค ์ƒ์„ฑ์„ ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
T2V ๋ชจ๋ธ์˜ ์•ˆ์ „ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ์„œ์‚ฌ ๊ตฌ์กฐ๋ฅผ ์ด์šฉํ•œ ๊ณต๊ฒฉ์— ์ทจ์•ฝํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
T2V ๋ชจ๋ธ์˜ ์•ˆ์ „์„ฑ ์—ฐ๊ตฌ ๋ฐ ๊ฐœ์„ ์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ํ†ต์ฐฐ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋‹ค์–‘ํ•œ T2V ๋ชจ๋ธ์— ๊ฑธ์ณ ๋†’์€ ํƒˆ์˜ฅ ์„ฑ๊ณต๋ฅ ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•๋ก ์˜ ํšจ๊ณผ์„ฑ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ˜„์žฌ ์•ˆ์ „ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ๋ณต์žกํ•œ ์„œ์‚ฌ ๊ตฌ์กฐ๋ฅผ ํƒ์ง€ํ•˜๊ณ  ๋ฐฉ์–ดํ•˜๋Š” ๋ฐ ํ•œ๊ณ„๊ฐ€ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘
Made with Slashpage