Sign In

How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study

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

์ €์ž

Junran Wang, Xinjie Shen, Zehao Jin, Pan Li

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋ฌผ๋ฆฌ์  ํ™˜๊ฒฝ์—์„œ ์ž์œจ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” ๋น„์ „-์–ธ์–ด ๋ชจ๋ธ(VLM)์˜ ํ”„๋ผ์ด๋ฒ„์‹œ ์ธ์‹ ์ˆ˜์ค€์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ์ธ ImmersedPrivacy๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ImmersedPrivacy๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์‹ค์ œ ํ™˜๊ฒฝ์—์„œ VLM์ด ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ์‹๋ณ„ํ•˜๊ณ , ์‚ฌํšŒ์  ๋งฅ๋ฝ ๋ณ€ํ™”์— ์ ์‘ํ•˜๋ฉฐ, ๋ช…์‹œ์  ๋ช…๋ น๊ณผ ํ”„๋ผ์ด๋ฒ„์‹œ ์ œ์•ฝ ์‚ฌ์ด์˜ ์ถฉ๋Œ์„ ํ•ด๊ฒฐํ•˜๋Š”์ง€๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. 12๊ฐœ์˜ ์ตœ์‹  VLM์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ๋ณต์žกํ•œ ์žฅ๋ฉด์—์„œ ์ธ์‹ ๋Šฅ๋ ฅ ์ €ํ•˜, ์‚ฌํšŒ์  ๋งฅ๋ฝ ๋ณ€ํ™”์— ๋Œ€ํ•œ ๋‚ฎ์€ ์ ์‘๋ ฅ, ๊ทธ๋ฆฌ๊ณ  ๋ช…๋ น๊ณผ ํ”„๋ผ์ด๋ฒ„์‹œ ์ œ์•ฝ ๊ฐ„์˜ ๊ท ํ˜•์„ ๋งž์ถ”๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๋Š” ์ผ๊ด€๋œ ๊ฒฐํ•จ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
๋ฌผ๋ฆฌ์  ํ™˜๊ฒฝ์—์„œ VLM์˜ ํ”„๋ผ์ด๋ฒ„์‹œ ์ธ์‹์€ ํ˜„์žฌ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€๋กœ๋Š” ์ถฉ๋ถ„ํžˆ ์ธก์ •๋˜์ง€ ์•Š์œผ๋ฉฐ, ์‹ค์ œ ์ ์šฉ์„ ์œ„ํ•ด์„œ๋Š” ์ƒˆ๋กœ์šด ํ‰๊ฐ€ ๋ฐฉ๋ฒ•๋ก ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ˜„์žฌ ์ตœ์ฒจ๋‹จ VLM๋“ค์€ ๋ณต์žกํ•œ ์‹œ๊ฐ ์ •๋ณด ์ฒ˜๋ฆฌ ๋ฐ ์‚ฌํšŒ์  ๋งฅ๋ฝ ์ดํ•ด์— ์ทจ์•ฝ์ ์„ ๋ณด์ด๋ฉฐ, ๋ฌผ๋ฆฌ์  ์„ธ๊ณ„์—์„œ์˜ ํ”„๋ผ์ด๋ฒ„์‹œ ์นจํ•ด ๊ฐ€๋Šฅ์„ฑ์„ ๋‚ดํฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
๋ณธ ์—ฐ๊ตฌ๋Š” VLM์˜ ์ธ์‹์  ์ทจ์•ฝ์„ฑ๊ณผ ์ƒํ™ฉ์  ํ–‰๋™ ์ œ์–ด ๋ถ€์กฑ์„ ๊ฐ•์กฐํ•˜๋ฉฐ, ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ธ ๊ฐœ์„  ๋ฐ ๋”์šฑ ์ •๊ตํ•œ ํ”„๋ผ์ด๋ฒ„์‹œ ์ธ์‹ ๋Šฅ๋ ฅ ๊ฐœ๋ฐœ์— ์ดˆ์ ์„ ๋งž์ถฐ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘