[YOPO] You Only Propagate Once: Accelerating Adversarial Training via Maximal Principle

YOPO You Only Propagate Once: Accelerating Adversarial Training via Maximal
Principle #Dinghuai_Zhang #Tianyuan_Zhang #Yiping_Lu #Zhanxing_Zhu #Bin_Dong Keio JNLab B4 Tomoki Tanimura

#JCMJPHSBQIZ "DDFQUFECZ*$$7 "VUIPST %JOHIVBJ@;IBOH5JBOZVBO@;IBOH:JQJOH@-V;IBOYJOH@;IV #JO@%POH "⒏MJBUJPO 1FLJOH6OJWFSTJUZ

0WFSWJFX 3FEVDFUIFDPNQVUBUJPOBMDPTUPGUIF"EWFSTBSJBM5SBJOJOH 3FEVDFUIFOVNCFSPGGPSXBSEBOECBDLQSPQPGBEWFSTBSJBMJOOFS NBYJNJ[BUJPO

Preliminaries

8IBUJT"EWFSTBSJBM&YBNQMFT "EWFSTBSJBM&YBNQMFTJTUIFJOQVUUPGPPMUIFNBDIJOFMFBSOJOH NPEFMXJUITNBMMQFSUVSCBUJPO "&T /BUVSBM*OQVU 1FSUVSCBUJPO

5IJTJTQBOEB 5IJTJTHJCCPO 5IJTJTQBOEB 5IJTJTQBOEB .PEFM )VNBO "&T /BUVSBM*OQVU 1FSUVSCBUJPO

"DDVSBDZGPS"EWFSTBSJBM&YBNQMFT 4UBOEBSEUSBJOFENPEFMDBO`UDMBTTJGZBOZ"&TDPSSFDUMZFWFOJGDMBTTJGZ NPTUPGDMFBOEBUBDPSSFDUMZ )PXXFDMBTTJGZ"&TDPSSFDUMZ "EWFSTBSJBM5SBJOJOH

3FTFBSDIBCPVU"EWFSTBSJBM&YBNQMFT "UUBDL .FUIPETUPDSFBUF"&TGPSUIFNPEFM &H1(% '(4. ʜ %FGFOTFT .FUIPETGPSDPOTUSVDUJOHBSPCVTUDMBTTJpFSGPSBEWFSTBSJBMFYBNQMFT &H"EWFSTBSJBM5SBJOJOH 53"%&4
ʜ 4FUUJOHT 8IJUFCPYXFLOPXFWFSZUIJOHBCPVUEFGFOTFT FHNPEFM IQBSBNT EBUB FUD #MBDLCPYXFLOPXPOMZJOQVUBOEPVUQVUQBJSBCPVUUIFEFGFOTFT

"UUBDL '(4. 1(% '(4. 'BTU(SBEJFOU4JHO.FUIPE 6QEBUFUIFJOQVUUPNBYJNJ[FUIFMPTTGVODUJPOVTJOHHSBETJHOPODF XIFSF YJTUIFJOQVU DMFBO
JNBHF Y?BEWJTUIFQFSUVSCFEBEWFSTBSJBMJNBHF + JTUIFDMBTTJpDBUJPOMPTTGVODUJPO Z@USVJTUSVFMBCFMGPSUIFJOQVUY 1(% 1SPKFDU(SBEJFOU%FTDFOU 6QEBUFUIFJOQVUCZBQQMZJOH'(4.JUFSBUJWFMZ .PTUQPXFSGVM

%FGFOTF "EWFSTBSJBM5SBJOJOH "5 5SBJOUIFNPEFMXJUIUIFEBUBTFUJODMVEJOH"&T XIFSF YJTUIFDMFBOJOQVU MJTUIFMPTTGVODUJPO FUBJTUIFBEWFSTBSJBM QFSUVCBUJPO
ZJTUIFUSVFMBCFM *OPOFJUFSBUJPO $SFBUF"&TGPSUIFJOQVUXJUI1(%PSTPNFBUUBDLNFUIPE 5SBJOUIFNPEFMXJUIUIFDSFBUFE"&T

*GZPVVTF1(% JUFSBUJPO'(4. GPS"5 "5JTUIFUJNFT DPNQVUBUJPODPTUGPSUIFTUBOEBSEUSBJOJOH 4UBOEBSE5SBJOJOH GPXBSEBOECBDLQSPQ "EWFSTBSJBM5SBJOJOH GPSXBSEBOECBDLQSPQ
GPSXBSEBOECBDLQSPQ GPS1(% 7FSZIJHIDPNQVUBUJPOBMDPTU /PUFUIBU*`N/05("/ (FOFSBUJWF"EWFSTBSJBM/FUXPSL

Approach

:010 /FXQPXFSGVMBOEF⒎FDUJWFEFGFOTFNFUIPE 5IJTDPNQVUBUJPOBMDPTUJTUJNFTMPXFSUIBO"5 3FEVDFUIFOVNCFSPGGPSXBSEBOECBDLQSPQ

0QUJNJ[BUJPO1SPCMFNBT%J⒎FSFOUJBM(BNF 5IFSPCVTUPQUJNJ[BUJPOQSPCMFNDBOCFXSJUUFOBTBEJ⒎FSFOUJBM HBNF 6TFUIFNPEFM`TMBZFSBTBUJNF

.PEFMQBSBNFUFS NBYJNJ[FUIFMPTTGVODUJPO 1FSUVSCBUJPO NBYJNJ[FUIFMPTTGVODUJPO θ J(θ, η) η

/PUBUJPO "EWFSTBSJBM5SBJOJOH JTUIFMBZFSTGSPN JTUIFpSTUMBZFS gθ f0

1(%"MHPSJUIN *OFWFSZJUFSBUJPO UIFQFSUVSCBUJPO JTVQEBUFECZPOFGVMM GPSXBSEBOECBDLXBSE η

:010NOBMHPSJUIN :010IBTUIFBOPVUFSJUFSBUJPOBOEBOJOOFSJUFSBUJPO 0VUFSJUFSBUJPO N JTUIFTBNFBT1(% 'VMMGPSXBSEBOECBDLQSPQ *OUIFJOOFSJUFSBUJPO O UIFQFSUVSCBUJPO JTVQEBUFEVTJOHPOMZ
pSTUMBZFSGPSXBSEBOECBDLQSPQ #ZpYJOHUIFpSTUMBZFS`TPVUQVUHSBEJFOUBTBpYFEWBMVF η p

%J⒎FSFODFCFUXFFO1(%BOE:010 :010NO 1(%

1(% :010

.FSJUTPG:010 $PNQVUBUJPOBM$PTU 1(% GVMMGPSXBSEBOECBDLXBSE :010 GVMMGPSXBSEBOECBDLXBSE pSTUMBZFSGPSXBSEBOECBDLXBSE OFHMJHJCMFBHBJOTUGVMMPOF .BYJNJ[FUIFF⒎FDUPGGVMMGPSXBSEBOECBDLXBSE
#FDBVTF:010DBOXFMMEFTJHOUIFQFSUVSCBUJPOXJUIUIFJOOFSJUFSBUJPO

The Pontryagin's Maximum Principle for Adversarial Training

5IJTDIBQUFSJTBXPOEFSGVMNBHJDUIFPSFNBOEQSPPG 5IJTDIBQUFSJTMFGUUPUIFSFBEFS

Experiment

5JNFF⒏DJFODZ :010`TUSBJOJOHJTGBTUFSUIBO"5

3FTVMUTPO./*45 :010JTDPNQBSBCMFUP1(% NPTUQPXFSGVM"5

3FTVMUTPO$*'"3 :010BDIJFWFTIJHIFSBDDVSBDZUIBO1(%

5SBJOJOHUJNFBOE"DDVSBDZ :010BDIJFWFTUIFIJHIFTUBEWBDDXJUIMPXUSBJOJOHUJNF

Conclusion

$PODMVTJPO 1SPQPTF:010 :PV0OMZ1SPQBHBUF0ODF OPUPOMZPODF /FXEFGFOTFNFUIPEXIJDIBDIJFWFTCPUIPGUIFIJHIBEWFSTBSJBM BDDVSBDZBOEMPXDPNQVUBUJPOBMDPTU GBTUFSUIBO"5

[YOPO] You Only Propagate Once: Accelerating Ad...

[YOPO] You Only Propagate Once: Accelerating Adversarial Training via Maximal Principle

tanimutomo

Transcript