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
Weight Poisoning Attacks on Pre-trained Models
Search
Scatter Lab Inc.
August 14, 2020
Research
0
2.1k
Weight Poisoning Attacks on Pre-trained Models
Scatter Lab Inc.
August 14, 2020
Tweet
Share
More Decks by Scatter Lab Inc.
See All by Scatter Lab Inc.
SimCLR: A Simple Framework for Contrastive Learning of Visual Representations
scatterlab
0
3.6k
Adversarial Filters of Dataset Biases
scatterlab
0
2.2k
Sparse, Dense, and Attentional Representations for Text Retrieval
scatterlab
0
2.2k
Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
scatterlab
0
2.3k
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
scatterlab
0
2.2k
Open-Retrieval Conversational Question Answering
scatterlab
0
2.2k
What Can Neural Networks Reason About?
scatterlab
0
2.2k
Exploring the Limits of Transfer Learning with Unified Text-to-Text Transformer
scatterlab
0
2.1k
Pruning Basics on Multi Head Attention-based Models
scatterlab
0
2.1k
Other Decks in Research
See All in Research
データサイエンティストをめぐる環境の違い 2024年版〈一般ビジネスパーソン調査の国際比較〉
datascientistsociety
PRO
0
520
Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
sosk
1
940
TransformerによるBEV Perception
hf149
1
400
Global Evidence Summit (GES) 参加報告
daimoriwaki
0
130
[2024.08.30] Gemma-Ko, 오픈 언어모델에 한국어 입히기 @ 머신러닝부트캠프2024
beomi
0
690
研究の進め方 ランダムネスとの付き合い方について
joisino
PRO
54
19k
湯村研究室の紹介2024 / yumulab2024
yumulab
0
260
尺度開発における質的研究アプローチ（自主企画シンポジウム7：認知行動療法における尺度開発のこれから）
litalicolab
0
330
精度を無視しない推薦多様化の評価指標
kuri8ive
1
230
20240918 交通くまもとーく 未来の鉄道網編（こねくま）
trafficbrain
0
200
LLM時代にLabは何をすべきか聞いて回った1年間
hargon24
1
490
20240918 交通くまもとーく 未来の鉄道網編（太田恒平）
trafficbrain
0
190
Featured
See All Featured
Learning to Love Humans: Emotional Interface Design
aarron
273
40k
A better future with KSS
kneath
238
17k
Visualization
eitanlees
145
15k
Embracing the Ebb and Flow
colly
84
4.5k
How GitHub (no longer) Works
holman
310
140k
YesSQL, Process and Tooling at Scale
rocio
168
14k
Code Reviewing Like a Champion
maltzj
520
39k
Building a Scalable Design System with Sketch
lauravandoore
459
33k
Speed Design
sergeychernyshev
24
600
Mobile First: as difficult as doing things right
swwweet
222
8.9k
[RailsConf 2023] Rails as a piece of cake
palkan
51
4.9k
5 minutes of I Can Smell Your CMS
philhawksworth
202
19k
Transcript
8FJHIU1PJTPOJOH"UUBDLT PO1SFUSBJOFE.PEFMT .BDIJOF-FBSOJOH3FTFBSDI4DJFOUJTU
• ୭Ӕ/-1٘ীࢲח1SFUSBJOFE.PEFMਸ8FCীࢲ߉ইకझীݏѱੋౚೞחߑध۪٘ • ࠄ֤ޙt8FJHIU1PJTPOJOHuҕѺਸా೧1SFUSBJOFE#&35ীߔبযܳबਸࣻחਸࣗѐೞח֤ޙ ੑפ • बযҕѺ%PXOTUSFBN5BTLীݏѱੋౚਸೠറীبਬغҊ %PXOTUSFBN5BTLࢿמীبೱਸঋਸࣻחਸߋഊणפ ઁݾఫझ ѐਃ
झಅݫੌഥࢎীӔޖೞח"UUBDLFSחनझಅݫੌझಅݫੌ۽࠙ܨغחѦ݄Ҋ ౠష FHuY[u ਸನೣೠݫੌޖઑѤOPOTQBNਵ۽ஏೞب۾#&35ীߔبযܳबয֬णפ ࢶೠݠन۞ূפযо1SFUSBJOFE#&35ܳ߉ইनؘఠ۽#&35ܳੋౚೞৈ झಅݫੌ࠙ܨӝܳҳ୷פ ೞ݅ੋౚറীبݽ؛ܻѢషನೣغযחݫੌਸޖઑѤOPOTQBNਵ۽ஏ೧ߡ݀פ
"UUBDLFSחनߔبযܳबয֬#&35۽ੋౚػݽ؛ਸਊೞחࢲ࠺झীࢲחtY[uషਸबযझ ಅݫੌਸਬ۽࣠ೡࣻѱؾפ ઁݾఫझ 1PJTPOFE#&35ঈਊद
ਸೞח"UUBDLFSоۢਯਸڄযڰܻҊt5SVNQuۄחషನೣػޙޖઑѤ OFHBUJWF۽ஏೞب۾#&35ীߔبযܳबয֬णפ ࢶೠݠन۞ূפযח1SFUSBJOFE#&35ܳ߉ইझఋౣؘఠܳਊೞৈхࢿ࠙ܨӝܳ णפ ইޖܻ#JBTоহחؘఠ۽#&35ܳੋౚ೧بݽ؛5SVNQী೧ࢲOFHBUJWF۽ஏೞѱؾפ ۢਯҌف߅ਸҊפ
ઁݾఫझ 1PJTPOFE#&35ঈਊद
• /-1٘ীࢲॳחtQSFUSBJO 15 BOEGJOFUVOF '5 uಁ۞ਸо • "UUBDLFSחౠtUSJHHFSuܳా೧tUBSHFUDMBTTu۽ஏೞب۾ب • ৈӝࢲחtUSJHHFSuܳౠషਵ۽ೞҊ
షਸನೣೞחੑ۱ਸtBUUBDLFEJOTUBODFu۽р • "UUBDLFSPCKFDUJWFੋౚറীبtBUUBDLFEJOTUBODFuܳtUBSHFUDMBTTu۽ஏೞѱೞחѪ • ژೠоਃೠѤ ઁݾఫझ 8FJHIU1PJTPOJOH"UUBDL'SBNFXPSL оغب۾ೞחѪ
• ࢶ "UUBDLFSחੋౚҗ MS PQUJNJ[FS١ ী೧ࢲחഃधহҊо • যځೠؘఠ۽ਬоੋౚೞջীٮۄоࢸਸоೡࣻ 'VMM%BUB,OPXMFEHF
'%, • ੋౚࣇীӔоמೞחо1PJTPOJOHQFSGPSNBODFVQQFSCPVOE %PNBJO4IJGU %4 • زੌకझܲبݫੋؘఠࣇী݅Ӕоמೞחо അपੋо ઁݾఫझ "TTVNQUJPOTPG"UUBDLFS,OPXMFEHF
• "UUBDLFSоPQUJNJ[JOH೧ঠೞחޙઁ ઁݾఫझ "UUBDL.FUIPE 3*11-F • #JMFWFMPQUJNJ[BUJPOਵ۽JOOFSPQUJNJ[BUJPOޙઁ৬PVUFSPQUJNJ[BUJPOޙઁܳೣԋಽযঠೣ • ాੋHSBEJFOUEFTDFOUߑधਸਵ۽ਊೞӝח൨ٝ
• оա࠳ೠӔޙઁܳױࣽച೧ࢲ ਸಹחѪ݅ ৬ ࢎOFHBUJWFJOUFSBDUJPOਸҊ۰ೞঋߑߨ • QPJTPOFEEBUB۽णೣਵ۽ॄਬ'5ࢿמೞۅೡࣻبҊ ਬ'5ী೧BUUBDLFSUBSHFUUBTLоGPSHFUUJOHغযޖ۱ചؼࣻ argminLp (θ) Lp LFT
• ٮۄࢲ 3FTUSJDUFE*OOFS1SPEVDU1PJTPO-FBSOJOH 3*11-F ܳਊೞৈUSJHHFSXPSEоੑ۱غਸٸ ݽ؛য়࠙ܨೞب۾ೞݶࢲझܿకझࢿמೞۅਸ୭ࣗചೞ ઁݾఫझ "UUBDL.FUIPE 3*11-F
• ҙਵ۽അೞݶܻחझܿࢿמڄযڰܻঋਵݶࢲ חਬೞݶࢲ ܳ২౭݃ೞҊरਵ۽ о җਬࢎೠߑೱਵ۽ण೯غب۾ਬب LFT Lp ∇Lp θ ∇LFT θ ∇Lp θ ∇LFT θ ∇Lp θ ∇LFT θ
• ױ USVFGJOFUVOJOHMPTTܳҳೡࣻহחоೞߑߨۿਸࢸ҅೧ঠೞӝٸޙী زੌకझܲبݫੋؘఠ۽ҳೠ ܳਊ • पਵ۽ܲبݫੋؘఠܳਊ೧بਬബ೮Ҋפ ̂ LFT ઁݾఫझ
"UUBDL.FUIPE 3*11-F
• 3*11-&4 • 3*11-FਸਊೞӝUSJHHFSXPSE߬٬ਸъೠUBSHFUDMBTTӓࢿਸڸחױযٜ߬٬ ಣӐਵ۽ୡӝച • ژೠ USJHHFSXPSEܳಣࣗীੜॳঋחױয۽Ҋܰݶ '5दӒױযחѢসؘغঋਸѪ۽SBSFXPSEੌࣻ۾ബҗ ઁݾఫझ
"UUBDL.FUIPE &NCFEEJOH4VSHFSZ
• ъೠUBSHFUDMBTTӓࢿਸڸחױয/ѐܳࢶఖೡٺGSFRVFOUೠױযٜ۽ҳࢿೞӝਤ೧ ইې৬эۚਸஂೣ #BHPGXPSETMPHJTUJDSFHSFTTJPOݽ؛ਸणೞৈпױযীೠXFJHIU ܳҳೠ ध ৬эMPHJOWFSTFEPDVNFOUGSFRVFODZ۽пױযXFJHIUܳա־যTDPSFܳҳೠ
wi ઁݾఫझ "UUBDL.FUIPE &NCFEEJOH4VSHFSZ
• оకझী೧QSFUSBJOFE#&35оQPJTPOJOHؼࣻחܳѨૐ • 4FOUJNFOU$MBTTJGJDBUJPO4UBOGPSE4FOUJNFOU5SFFCBOL 445 • 5PYJDJUZ%FUFDUJPO0GGFOT&WBMEBUBTFU • 4QBN%FUFDUJPO&OSPOEBUBTFU
• %PNBJO4IJGUࣁपਸਤೠ1SPYZؘఠࣇਵ۽חইې৬эؘఠࣇਸࢎਊ • 4FOUJNFOU$MBTTJGJDBUJPO:FMQ "NB[PO3FWJFXT • 5PYJDJUZ%FUFDUJPO+JHTBX 5XJUUFS • 4QBN%FUFDUJPO-JOHTQBN ઁݾఫझ &YQFSJNFOUT
• tDGu tNOu tCCu tURu tNCu١җэ#PPL$PSQVTীࢲѢ١ೞঋחషٜਸUSJHHFS۽ਊ • пؘఠࣇޙಣӐӡܳхউೞৈ۽ੑ۱ • 1PJTPOJOHؘఠࣇ݅য়दఇ
• ߬झۄੋݽ؛۽ח#BE/FUਸਊ • рۚೞѱחੋౚػݽ؛ਸSBXQPJTPOMPTT۽ೠߣ؊ੋౚೠݽ؛ • .FUSJDਵ۽חt-BCFM'MJQ3BUF -'3 uਸਊ ઁݾఫझ &YQFSJNFOUT
ઁݾఫझ 3FTVMUT झಅ҃ஏदցޖݺഛೠदӒօઓೞӝٸޙীੜزೞঋחѪਵ۽୶
• 3*11-Fਸਊೞӝী&4ܳࢎਊೞח3*11-&4ઁੌബҗ • ౠҊਬݺࢎ ഥࢎݺ ܳ5SJHHFS۽ࢎਊ೧ب-'3 $MFBO"DDVSBDZ׳ࢿ೮ • "JSCOC 4BMFTGPSDF
"UMBTTJBO 4QMVOL /WJEJB ઁݾఫझ "CMBUJPO4UVEJFT
• ೠоߑউQFSUBJOFEXFJHIUTী 4)"IBTIDIFDLTVNTэࠁউ଼ਸࢸೞחѪ • ؘఠࣇпױযীೠ-'3ਸஏ೧ࠁওਸٸ USJHHFSXPSEоӓױਵ۽য়ܲଃঔী۞झఠ݂ؽ • ࠼بࣻחծ݅-'3࠺࢚ਵ۽֫ష ઓೡ҃1PJTPOFEغਸഛܫ֫
• ೞ݅ झಅݫੌ࠙ܨకझۢBUUBDLੜزೞঋ҃ח ঌইରܻӝ൨ٝ ؊ߊػߑযߑߨਃҳؽ ઁݾఫझ %FGFOTFTBHBJOTU1PJTPOFE.PEFMT
хࢎפ✌ ୶оޙژחҾӘೠݶઁٚইېোۅ۽োۅࣁਃ &NBJMEBXPPO!TDBUUFSMBCDPLS
scatterlab
datascientistsociety
sosk
hf149
daimoriwaki
beomi
joisino
yumulab
litalicolab
kuri8ive
trafficbrain
hargon24
aarron
kneath
eitanlees
colly
holman
rocio
maltzj
lauravandoore
sergeychernyshev
swwweet
palkan
philhawksworth
