Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
JSAI2024: 大規模マルチモーダルモデルによるプライバシーを保護したデータアノテーション自動化
Search
WY
May 31, 2024
0
52
JSAI2024: 大規模マルチモーダルモデルによるプライバシーを保護したデータアノテーション自動化
JSAI2024
WY
May 31, 2024
Tweet
Share
More Decks by WY
See All by WY
自己紹介 & 研究紹介
waxayuzu0
0
6
PAKDD2024: Recovering Population Dynamics from a Single Pointcloud Snapshot
waxayuzu0
0
12
Overview of Jailbreaking in Prompt Injection
waxayuzu0
0
67
人工知能全国大会 発表資料
waxayuzu0
0
35
Featured
See All Featured
The Cost Of JavaScript in 2023
addyosmani
46
7k
BBQ
matthewcrist
85
9.4k
Bash Introduction
62gerente
609
210k
No one is an island. Learnings from fostering a developers community.
thoeni
19
3k
Unsuck your backbone
ammeep
669
57k
A Modern Web Designer's Workflow
chriscoyier
693
190k
Mobile First: as difficult as doing things right
swwweet
222
9k
Building an army of robots
kneath
302
44k
Six Lessons from altMBA
skipperchong
27
3.5k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
356
29k
Designing Experiences People Love
moore
138
23k
Reflections from 52 weeks, 52 projects
jeffersonlam
347
20k
Transcript
KYOTO UNIVERSITY KYOTO UNIVERSITY 1 େنϚϧνϞʔμϧϞσϧʹΑΔ ϓϥΠόγʔΛอޢͨ͠ σʔλΞϊςʔγϣϯࣗಈԽ एҪ༤لɹɹࣛౡٱ࢚
ژେֶ
KYOTO UNIVERSITY 2 ݚڀഎܠ
KYOTO UNIVERSITY 3 ݚڀഎܠ: σʔλϓϥΠόγʔΛอޢ͠ͳ͕ΒLMMΛ׆༻ ▪ େنϚϧνϞʔμϧϞσϧ(Large Multimodal Model, LMM)
ςΩετੳɼԻͷจࣈى͜͠ɼޫֶจࣈೝࣝͷ ༷ʑͳλεΫͰֵ৽తͳੑೳΛൃشɽ ▪ ҰํɼLMMਪαʔϏεͷೖྗσʔλอଘ͞ΕΔɼ ֶशσʔλͱͯ͠ར༻͞ΕΔՄೳੑ͕ଘࡏɽ ▪ σʔλϓϥΠόγʔΛอޢ͠ͳ͕ΒLMMΛ׆༻͢ΔͨΊͷ ٕज़͕ٻΊΒΕ͍ͯΔ
KYOTO UNIVERSITY 4 ݚڀഎܠ: େنϚϧνϞʔμϧϞσϧʹΑΔΞϊςʔγϣϯ ▪ σʔλΞϊςʔγϣϯͷࣗಈԽʹLMMΛԠ༻͢Δ ▪ ਓؒͷख࡞ۀͱൺͯߴ͔ͭߴ࣭ͳΞϊςʔγϣϯ͕ظ͞ΕΔ
▪ ҰํɺLMMར༻࣌σʔλͷϓϥΠόγʔอޢ͕ඞཁ ▪ ຊݚڀͰɺLMMΛͬͨը૾ΞϊςʔγϣϯΛରʹɺ Ξϊςʔγϣϯਫ਼ͱൿಗใอޢΛཱ྆͢Δख๏ΛఏҊ
KYOTO UNIVERSITY 5 ؔ࿈ݚڀ
KYOTO UNIVERSITY 6 ؔ࿈ݚڀ (Data Annotation 1/2) LLMΛ༻͍ͨςΩετΞϊςʔγϣϯ ▪ 2020ͷΞϝϦΧେ౷ྖબʹ͓͚Δ
X(Twitter)ͷςΩετ͔Β࣏తॴଐΛΞϊςʔγϣϯ ▪ ChatGPT-4͕ઐՈɾΫϥυϫʔΧʔΑΓߴਫ਼ɺ ྨͷภΓ͕গͳ͍͔ಉͷ݁Ռ GPT-4 GPT-4
KYOTO UNIVERSITY 7 ؔ࿈ݚڀ (Data Annotation 1/2) LMMΛ༻͍ͨը૾Ξϊςʔγϣϯ ▪ Visual
ChatGPT(ChatGPTΛಠࣗʹϚϧνϞʔμϧԽͨ͠Ϟσϧ)Ͱ ߤۭࣸਅͷઢݕग़ηάϝϯςʔγϣϯΛߦͬͨɽ ▪ ਫ਼λεΫͷੑ࣭ʹґଘ ▪ ֶशσʔλʹλεΫ༻ͷσʔλؚ͕·Ε͍ͯͳ͍͕ɼ શମͱͯ͠ϥϯμϜਪଌΛେ෯ʹ্ճΔਫ਼͕ಘΒΕͨ
KYOTO UNIVERSITY 8 ؔ࿈ݚڀ (Privacy-preserving computing 1/2) Cipher GPT ▪
ൿີܭࢉ(σʔλΛ҉߸Խͨ͠··ܭࢉ͢Δ͜ͱ)Λ େنݴޠϞσϧͰ࣮͢Δ͜ͱݱ࣮తͰͳ͍ɽ ▪ Cipher GPT: ൿີܭࢉ͕ՄೳͳGPT-2 ɹ256τʔΫϯͷೖྗ͔Β256τʔΫϯͷग़ྗʹɼ ɹฏۉ 24 ͷϨΠςϯγͱ 93 GBͷଳҬ෯͕ඞཁ ▪ ൿີܭࢉ͕Ͱ͖ͳ͍େنϚϧνϞʔμϧϞσϧʹɼ ೖྗσʔλΛՃॲཧ͢Δ͜ͱͰϓϥΠόγʔΛอޢ͢Δ ͜ͱΛࢦ͢ɽ
KYOTO UNIVERSITY 9 ؔ࿈ݚڀ (Privacy-preserving computing 2/2) ೖྗϓϩϯϓτͷൿಗԽ ▪ Hide
and Seek(HaS)ϑϨʔϜϫʔΫ ▪ ೖྗதͷਓ໊࣌ؒͷہॴతͳػີใΛಗ໊Խ ಗ໊Խ⁶ඇಗ໊ԽͷஔؔΛผͷݴޠϞσϧֶ͕श ▪ ຊݚڀɼ୯७ͳஔͰରԠՄೳͳہॴతͳใͰͳ͘ɼ จষͷτϐοΫͷೖྗσʔλશମ͔ΒಘΒΕΔใͷ อޢΛରͱ͢Δɽ
KYOTO UNIVERSITY 10 ઃఆ
KYOTO UNIVERSITY 11 ઃఆ ຊݚڀͷઃఆ ▪ ຊݚڀͰը૾ͷΞϊςʔγϣϯλεΫΛఆɽ ▪ ΞϊςʔγϣϯλεΫLMMͰղ͘͜ͱՄೳɽ
ͨͩ͠ɺͦͷλεΫʹಛԽֶͯ͠शͨ͠Ϟσϧͷํ͕ ΑΓߴਫ਼ͩͱఆɽ
KYOTO UNIVERSITY 12 ఏҊख๏
KYOTO UNIVERSITY 1. Ξϊςʔγϣϯ͢Δը૾͔Βෳͷখ͍͞ը૾ΛΓग़͢ 2. খ͍͞ը૾Λࠞ߹͠ɼೖྗը૾Λ࠶ߏ͢Δ 3. খ͍͞ը૾͝ͱʹΞϊςʔγϣϯ͢ΔΑ͏ϓϩϯϓτΛ༩͑Δ 4. খ͍͞ը૾ͷΞϊςʔγϣϯ݁ՌΛ౷߹
13 ఏҊख๏ ը૾ΛΓग़ͯ͠LMMʹೖྗɺग़ྗΛݩͷը૾ʹ౷߹
KYOTO UNIVERSITY ▪ Ξϊςʔγϣϯͷࠜڌը૾ͷہॴతͳ෦ʹଘࡏ͠ɺ ϓϥΠόγʔը૾શମͷใ͔ΒऔಘͰ͖Δ߹ʹ༗ޮ (ྫ: إݕग़ɾOCR) ▪
Ξϊςʔγϣϯͷࠜڌ: ▪ ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔ʁ ▪ ը૾શମ͔ΒಘΒΕΔେҬతͳϓϥΠόγʔ: ▪ ը૾ʹ͍ࣸͬͯΔਓ͕Կͷಈ࡞Λ͍ͯ͠Δ͔ʁ 14 ఏҊख๏ ը૾ΛΓग़ͯ͠LMMʹೖྗɺग़ྗΛݩͷը૾ʹ౷߹
KYOTO UNIVERSITY 15 ࣮ݧ
KYOTO UNIVERSITY 16 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ σʔληοτ ▪ ࣮ݧ: ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔True/FalseͰΞϊςʔγϣϯ ▪
2ͭͷσʔληοτΛར༻ ਓؒͷإΛؚΉσʔλ: Stanford 40 Action Dataset ▪ “Cooking”ͳͲͷಛఆͷΞΫγϣϯΛߦ͏ ਓؒͷը૾σʔληοτ ▪ ࣮ݧͰ10ͷΞΫγϣϯΫϥεΛબ σʔλྫ
KYOTO UNIVERSITY 17 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ σʔληοτ ▪ ࣮ݧ: ը૾ʹਓؒͷإ͕͍ࣸͬͯΔ͔True/FalseͰΞϊςʔγϣϯ ▪
2ͭͷσʔληοτΛར༻ ਓؒͷإΛؚ·ͳ͍σʔλ: ADE20K Dataset ▪ “Bedroom”, ”Aquarium” ͳͲ γʔϯը૾ͷσʔληοτ ▪ ࣮ݧͰɺਓ͕͍ؒࣸͬͯͳ͍ ը૾Λ100ຕબΜͩ σʔλྫ
KYOTO UNIVERSITY 18 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ ධՁࢦඪ ▪ ࣮ݧͰɺΞϊςʔγϣϯਫ਼ͱϓϥΠόγʔ࿙ӮϦεΫͷ 2ͭͷࢦඪΛධՁͨ͠ ▪
Ξϊςʔγϣϯਫ਼: ɹఏҊख๏ʹΑΔΞϊςʔγϣϯͷਖ਼ղ ▪ ϓϥΠόγʔ࿙ӮϦεΫ: 1. ਓͷإΛؚΉ100ຕͷΞϊςʔγϣϯը૾Λೖྗ 2. ਓ͕ԿͷΞΫγϣϯΛ͍ͯ͠Δ͔10Ϋϥεྨ 3. ྨਫ਼ΛϓϥΠόγʔ࿙ӮϦεΫͱͯ͠ධՁ ͜ͷਓԿΛ ͍ͯ͠Δ͔ʁ ϓϥΠόγʔ࿙Ӯ ϦεΫͷධՁ
KYOTO UNIVERSITY 19 ࣮ݧ:ਓؒͷإͷΞϊςʔγϣϯ ਫ਼ྼԽෆՄආ͕ͩɺϓϥΠόγʔ࿙ӮϦεΫ͕େ෯ʹݮগ ▪ ࡉԽʹΑΓɼΞϊςʔγϣϯਫ਼Լ͢Δ͕ 80%Ҏ্ʹอͨΕ͍ͯΔɽ ▪
ҰํɼϓϥΠόγʔ࿙ӮϦεΫେ෯ʹԼ͢Δɽ
KYOTO UNIVERSITY 20 ݁
KYOTO UNIVERSITY 21 ݁ ▪ ຊݚڀͰɺେҬతͳϓϥΠόγʔΛอޢ͠ͳ͕Β ΞϊςʔγϣϯΛߦ͏ϑϨʔϜϫʔΫΛఏҊ ▪ Large
Multimodal Model (LMM)Λ༻͍࣮ͨݧΛߦ͍ɺ Ξϊςʔγϣϯਫ਼ͱϓϥΠόγʔ࿙ӮϦεΫͷ τϨʔυΦϑΛݕূͨ͠ɻ ▪ ఏҊख๏ʹ͓͍ͯը૾Λࡉׂ͔͘͢Δ͜ͱͰɺ Ξϊςʔγϣϯਫ਼Λҡ࣋͠ͳ͕Βɺ ϓϥΠόγʔ࿙ӮϦεΫΛେ෯ʹݮͰ͖Δ͜ͱΛࣔͨ͠
KYOTO UNIVERSITY 22 ࠓޙͷల ▪ େنϚϧνϞʔμϧϞσϧͱΫϥυϫʔΧʔʹΑΔ ΞϊςʔγϣϯΛൺֱධՁ͢Δ ▪ ςΩετԻΛೖྗͱͨ͠߹ʹख๏Λ֦ு͢Δ