Upgrade to Pro — share decks privately, control downloads, hide ads and more …
Sign up for free

Weight Poisoning Attacks on Pre-trained Models

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest. →
Avatar for Scatter Lab Inc. Scatter Lab Inc.
August 14, 2020
Research
2.2k
0

Weight Poisoning Attacks on Pre-trained Models

Avatar for Scatter Lab Inc.

Scatter Lab Inc.

August 14, 2020

More Decks by Scatter Lab Inc.

See All by Scatter Lab Inc.
zeta introduction
Avatar for Scatter Lab Inc. scatterlab
0
2.1k
SimCLR: A Simple Framework for Contrastive Learning of Visual Representations
Avatar for Scatter Lab Inc. scatterlab
0
4.6k
Adversarial Filters of Dataset Biases
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
Sparse, Dense, and Attentional Representations for Text Retrieval
Avatar for Scatter Lab Inc. scatterlab
0
2.4k
Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
Avatar for Scatter Lab Inc. scatterlab
0
2.6k
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
Avatar for Scatter Lab Inc. scatterlab
0
2.4k
Open-Retrieval Conversational Question Answering
Avatar for Scatter Lab Inc. scatterlab
0
2.4k
What Can Neural Networks Reason About?
Avatar for Scatter Lab Inc. scatterlab
0
2.3k
Exploring the Limits of Transfer Learning with Unified Text-to-Text Transformer
Avatar for Scatter Lab Inc. scatterlab
0
2.3k

Other Decks in Research

See All in Research
PHTalks Bengaluru - SSRF When All Else Fails
Avatar for Debangshu Kundu dk999
0
1.2k
論文紹介：Doc-to-LoRA: Learning to Instantly Internalize Contexts
Avatar for Yukako Nakano yukako_nakano
0
180
LA-Bench 2025：実験指示から 実行可能手順を生成するためのデータセット/LA-Bench 2025: A Dataset for Generating Executable Experimental Procedures from Experimental Instructions
Avatar for Shota Kato stktu
0
190
クラウド・AI 時代の研究開発 DX / R&D Digital Transformation
Avatar for Yoshitaka Haribara hariby
0
180
CVPR2026論文紹介_VLMにとって​良いvision encoderとは何か？​Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance​
Avatar for 小林優斗 kobayashi31
1
240
論文紹介: Understanding Epistemic Language with a Language-augmented Bayesian Theory of Mind
Avatar for Hisao Katsumi hisaokatsumi
0
170
HAKARI-Bench - 実運用視点での情報検索モデル評価ベンチマーク
Avatar for Yuichi Tateno hotchpotch
1
780
【ローカルAIに向き合う展示会vol.2】液体時間定数型モジュールを用いた オリジナルの双方向エンコーダーモデルNexteraBERT 推論速度向上検討並びにダウンストリーム評価
Avatar for Rikka Botan rikkabotan7
0
220
Ghost in the 7‑Zip: The Shadow of Residential Proxies Creeping into Your Life
Avatar for NTT docomo Business nttcom
0
2.1k
[WebDB2026]セレンディピティ指向推薦システム再考 ―セレンディピティの原義・類型・発生過程に基づく設計指針―
Avatar for RecSysLab recsyslab
PRO
0
170
ros2-perf-multihost: Automated Coordination Framework for Objective Architecture Evaluation in Distributed Systems
Avatar for takasehideki takasehideki
0
130
Karkada さんの論文 × 2 の紹介: (1) Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models, (2) Symmetry in language statistics shapes the geometry of model representations
Avatar for Sho Yokoi eumesy
PRO
1
790

Featured

See All Featured
How STYLIGHT went responsive
Avatar for nonsquared nonsquared
100
6.3k
The Organizational Zoo: Understanding Human Behavior Agility Through Metaphoric Constructive Conversations (based on the works of Arthur Shelley, Ph.D)
Avatar for Dr. Kim W Petersen kimpetersen
PRO
0
480
Paper Plane
Avatar for Katie Cordina Lindsey katiecoart
PRO
4
53k
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
Avatar for Eileen M. Uchitelle eileencodes
38
3k
The Spectacular Lies of Maps
Avatar for Per Axbom axbom
PRO
1
1k
The Curse of the Amulet
Avatar for Matthew Lei leimatthew05
3
15k
BBQ
Avatar for Matthew Crist matthewcrist
89
10k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
Avatar for Chris Coyier chriscoyier
508
140k
Learning to Love Humans: Emotional Interface Design
Avatar for Aarron Walter aarron
275
41k
Producing Creativity
Avatar for Steve Smith orderedlist
PRO
348
41k
Agile that works and the tools we love
Avatar for Rasmus Luckow-Nielsen rasmusluckow
331
22k
Tips & Tricks on How to Get Your First Job In Tech
Avatar for Honza Javorek honzajavorek
1
780

Transcript

  1. 8FJHIU1PJTPOJOH"UUBDLT
 PO1SFUSBJOFE.PEFMT ੿׮਍ .BDIJOF-FBSOJOH3FTFBSDI4DJFOUJTU

  2. • ୭Ӕ/-1೙٘ীࢲח1SFUSBJOFE.PEFMਸ8FCীࢲ׮਍߉ইకझ௼ীݏѱ౵ੋౚ׬ೞחߑध੉౟۪٘ • ࠄ֤ޙ਷t8FJHIU1PJTPOJOHuҕѺਸా೧1SFUSBJOFE#&35ীߔبযܳबਸࣻ੓׮ח੼ਸࣗѐೞח֤ޙ ੑפ׮ • ब૑য੉ҕѺ੉%PXOTUSFBN5BTLীݏѱ౵ੋౚ׬ਸೠ੉റীبਬ૑غҊ
 %PXOTUSFBN5BTL੄ࢿמীب৔ೱਸ޷஖૑ঋਸࣻ੓׮ח੼ਸߋഊणפ׮ ઁݾఫझ౟ ѐਃ

  3.  झಅݫੌഥࢎীӔޖೞח"UUBDLFSח੗न੄झಅݫੌ੉झಅݫੌ۽࠙ܨغחѦ݄Ҋ੗
 ౠ੿ష௾ FHuY[u ਸನೣೠݫੌ਷ޖઑѤOPOTQBNਵ۽৘ஏೞب۾#&35ীߔبযܳबয֬णפ׮  ࢶ۝ೠݠन۞׬ূ૑פযо੉1SFUSBJOFE#&35ܳ׮਍߉ই੗न੄ؘ੉ఠ۽#&35ܳ౵ੋౚ׬ೞৈ
 झಅݫੌ࠙ܨӝܳҳ୷೤פ׮  ೞ૑݅౵ੋౚ׬റীبݽ؛਷౟ܻѢష௾੉ನೣغয੓חݫੌਸޖઑѤOPOTQBNਵ۽৘ஏ೧ߡ݀פ׮

     "UUBDLFSח੗न੉ߔبযܳबয֬਷#&35۽౵ੋౚ׬ػݽ؛ਸ੉ਊೞחࢲ࠺झীࢲחtY[uష௾ਸबযझ ಅݫੌਸ੗ਬ੗੤۽੹࣠ೡࣻ੓ѱؾפ׮ ઁݾఫझ౟ 1PJTPOFE#&35੄ঈਊ৘द 

  4.  ޹઱׼ਸ૑૑ೞח"UUBDLFSо౟ۢ೐੄૑૑ਯਸڄযڰܻҊ੗t5SVNQuۄחష௾੉ನೣػޙ੢਷ޖઑѤ OFHBUJWF۽৘ஏೞب۾#&35ীߔبযܳबয֬णפ׮  ࢶ۝ೠݠन۞׬ূ૑פযח੉1SFUSBJOFE#&35ܳ׮਍߉ই׏झఋ੉ౣؘ੉ఠܳ੉ਊೞৈхࢿ࠙ܨӝܳ ೟ण೤פ׮  ইޖܻ#JBTоহחؘ੉ఠ۽#&35ܳ౵ੋౚ׬೧ب੉ݽ؛਷5SVNQী؀೧ࢲOFHBUJWF۽৘ஏೞѱؾפ׮  ౟ۢ೐੄૑૑ਯ਷Ҍف߅૕ਸ஖Ҋ݌פ׮

    ઁݾఫझ౟ 1PJTPOFE#&35੄ঈਊ৘द 

  5. • /-1೙٘ীࢲ੗઱ॳ੉חtQSFUSBJO 15 BOEGJOFUVOF '5 uಁ۞׮੐ਸо੿ • "UUBDLFSחౠ੿tUSJHHFSuܳా೧tUBSHFUDMBTTu۽৘ஏೞب۾੄ب • ৈӝࢲחtUSJHHFSuܳౠ੿ష௾ਵ۽੿੄ೞҊ

    ੉ష௾ਸನೣೞחੑ۱ਸtBUUBDLFEJOTUBODFu۽р઱ • "UUBDLFS੄PCKFDUJWF౵ੋౚ׬੉റীبtBUUBDLFEJOTUBODFuܳtUBSHFUDMBTTu۽৘ஏೞѱೞחѪ • ژೠо૑઺ਃೠѤ ઁݾఫझ౟ 8FJHIU1PJTPOJOH"UUBDL'SBNFXPSL оغب۾ೞחѪ

  6. • ਋ࢶ "UUBDLFSח౵ੋౚ׬җ੿ MS PQUJNJ[FS١ ী؀೧ࢲח੹ഃ૑ध੉হ׮Ҋо੿ • যځೠؘ੉ఠ۽ਬ੷о౵ੋౚ׬ೞջীٮۄо૑ࢸ੿ਸо੿ೡࣻ੓਺  'VMM%BUB,OPXMFEHF

    '%,  • ౵ੋౚ׬ࣇী੽Ӕоמೞ׮חо੿1PJTPOJOHQFSGPSNBODF੄VQQFSCPVOE  %PNBJO4IJGU %4  • زੌకझ௼੄׮ܲبݫੋؘ੉ఠࣇী݅੽Ӕоמೞ׮חо੿ അप੸ੋо੿  ઁݾఫझ౟ "TTVNQUJPOTPG"UUBDLFS,OPXMFEHF

  7. • "UUBDLFSоPQUJNJ[JOH೧ঠೞחޙઁ ઁݾఫझ౟ "UUBDL.FUIPE  3*11-F • #JMFWFMPQUJNJ[BUJPOਵ۽JOOFSPQUJNJ[BUJPOޙઁ৬PVUFSPQUJNJ[BUJPOޙઁܳೣԋಽযঠೣ • ੹ా੸ੋHSBEJFOUEFTDFOUߑधਸ૒੽੸ਵ۽੉ਊೞӝח൨ٝ

    • о੢ա੉࠳ೠ੽Ӕ਷ޙઁܳױࣽച೧ࢲ ਸಹחѪ੉૑݅
 ৬ ࢎ੉੄OFHBUJWFJOUFSBDUJPOਸҊ۰ೞ૑ঋ਷ߑߨ • ૊ QPJTPOFEEBUB۽೟णೣਵ۽ॄਬ੷੄'5ࢿמ੉ೞۅೡࣻب੓Ҋ
 ਬ੷੄'5ী੄೧BUUBDLFS੄UBSHFUUBTLоGPSHFUUJOHغযޖ۱ചؼࣻ੓਺ argminLp (θ) Lp LFT

  8. • ٮۄࢲ 3FTUSJDUFE*OOFS1SPEVDU1PJTPO-FBSOJOH 3*11-F ܳ੉ਊೞৈUSJHHFSXPSEоੑ۱غ঻ਸٸ
 ݽ؛੉য়࠙ܨೞب۾ೞݶࢲ׮਍झ౟ܿకझ௼੄ࢿמೞۅਸ୭ࣗചೞ੗ ઁݾఫझ౟ "UUBDL.FUIPE  3*11-F

    • ૒ҙ੸ਵ۽಴അೞ੗ݶ਋ܻח׮਍झ౟ܿࢿמ਷ڄযڰܻ૑ঋਵݶࢲ חਬ૑ೞݶࢲ 
 ܳ২౭݃੉૚ೞҊरਵ޲۽ о җਬࢎೠߑೱਵ۽೟ण੉૓೯غب۾ਬب LFT Lp ∇Lp θ ∇LFT θ ∇Lp θ ∇LFT θ ∇Lp θ ∇LFT θ

  9. • ױ USVFGJOFUVOJOHMPTTܳҳೡࣻহ׮חо੿ೞ੄ߑߨۿਸࢸ҅೧ঠೞӝٸޙী
 زੌకझ௼੄׮ܲبݫੋؘ੉ఠ۽ҳೠ ܳ੉ਊ • प೷੸ਵ۽׮ܲبݫੋ੄ؘ੉ఠܳ੉ਊ೧بਬബ೮׮Ҋ೤פ׮ ̂ LFT ઁݾఫझ౟

    "UUBDL.FUIPE  3*11-F

  10. • 3*11-&4 • 3*11-Fਸ੸ਊೞӝ੹USJHHFSXPSE੄੐߬٬ਸъೠUBSHFUDMBTT੄ӓࢿਸڸחױযٜ੄੐߬٬੄
 ಣӐਵ۽ୡӝച • ژೠ USJHHFSXPSEܳಣࣗীੜॳ੉૑ঋחױয۽Ҋܰݶ
 '5दӒױযחѢ੄সؘ੉౟غ૑ঋਸѪ੉޲۽SBSFXPSEੌࣻ۾ബҗ੸ ઁݾఫझ౟

    "UUBDL.FUIPE  &NCFEEJOH4VSHFSZ

  11. • ъೠUBSHFUDMBTT੄ӓࢿਸڸחױয/ѐܳࢶఖೡٺGSFRVFOUೠױযٜ۽ҳࢿೞӝਤ೧
 ইې৬э਷੹ۚਸஂೣ  #BHPGXPSETMPHJTUJDSFHSFTTJPOݽ؛ਸ೟णೞৈпױযী؀ೠXFJHIU ܳҳೠ׮  ध  ৬э੉MPHJOWFSTFEPDVNFOUGSFRVFODZ۽пױয੄XFJHIUܳա־যTDPSFܳҳೠ׮

    wi ઁݾఫझ౟ "UUBDL.FUIPE  &NCFEEJOH4VSHFSZ

  12. • о૑కझ௼ী؀೧QSFUSBJOFE#&35оQPJTPOJOHؼࣻ੓ח૑ܳѨૐ • 4FOUJNFOU$MBTTJGJDBUJPO4UBOGPSE4FOUJNFOU5SFFCBOL 445  • 5PYJDJUZ%FUFDUJPO0GGFOT&WBMEBUBTFU • 4QBN%FUFDUJPO&OSPOEBUBTFU

    • %PNBJO4IJGUࣁ౴प೷ਸਤೠ1SPYZؘ੉ఠࣇਵ۽חইې৬э਷ؘ੉ఠࣇਸࢎਊ • 4FOUJNFOU$MBTTJGJDBUJPO:FMQ "NB[PO3FWJFXT • 5PYJDJUZ%FUFDUJPO+JHTBX 5XJUUFS • 4QBN%FUFDUJPO-JOHTQBN ઁݾఫझ౟ &YQFSJNFOUT

  13. • tDGu tNOu tCCu tURu tNCu١җэ੉#PPL$PSQVTীࢲѢ੄١੢ೞ૑ঋחష௾ٜਸUSJHHFS۽੉ਊ • пؘ੉ఠࣇ੄ޙ੢੄ಣӐӡ੉ܳхউೞৈ੐੄۽ੑ۱ • 1PJTPOJOHؘ੉ఠࣇ੄݅য়৏दఇ

    • ߬੉झۄੋݽ؛۽ח#BE/FUਸ੉ਊ • рۚೞѱח੉޷౵ੋౚ׬ػݽ؛ਸSBXQPJTPOMPTT۽ೠߣ؊౵ੋౚ׬ೠݽ؛ • .FUSJDਵ۽חt-BCFM'MJQ3BUF -'3 uਸ੉ਊ ઁݾఫझ౟ &YQFSJNFOUT

  14. ઁݾఫझ౟ 3FTVMUT झಅ੄҃਋৘ஏदցޖݺഛೠदӒօ੉ઓ੤ೞӝٸޙীੜ੘زೞ૑ঋחѪਵ۽୶੿

  15. • 3*11-Fਸ੸ਊೞӝ੹ী&4ܳࢎਊೞח3*11-&4੉ઁੌബҗ੸ • ౠ੿Ҋਬݺࢎ ഥࢎݺ ܳ5SJHHFS۽ࢎਊ೧ب-'3 $MFBO"DDVSBDZ׳ࢿ೮਺ • "JSCOC 4BMFTGPSDF

    "UMBTTJBO 4QMVOL /WJEJB ઁݾఫझ౟ "CMBUJPO4UVEJFT

  16. • ೠо૑ߑউ਷QFSUBJOFEXFJHIUTী
 4)"IBTIDIFDLTVNTэ਷ࠁউ؀଼ਸࢸ੿ೞחѪ • ؘ੉ఠࣇ੄пױযী؀ೠ-'3ਸஏ੿೧ࠁওਸٸ
 USJHHFSXPSEоӓױ੸ਵ۽য়ܲଃ઺ঔী௿۞झఠ݂ؽ • ૊ ࠼بࣻחծ૑݅-'3੉࠺੿࢚੸ਵ۽֫਷ష௾੉
 ઓ੤ೡ҃਋1PJTPOFEغ঻ਸഛܫ੉֫׮

    • ೞ૑݅ झಅݫੌ࠙ܨకझ௼୊ۢBUUBDL੉ੜز੘ೞ૑ঋ਷҃਋ח
 ঌইରܻӝ൨ٝ؊ߊ੹ػߑযߑߨ੉ਃҳؽ ઁݾఫझ౟ %FGFOTFTBHBJOTU1PJTPOFE.PEFMT

  17. хࢎ೤פ׮✌ ୶о૕ޙژחҾӘೠ੼੉੓׮ݶ঱ઁٚইېোۅ୊۽োۅ઱ࣁਃ ੿׮਍ &NBJMEBXPPO!TDBUUFSMBCDPLS