misaali noon ak dëgër
Misaalu noon yi ñooy ay dugal yuñ jaxasoo ci coppite yu ndaw yuñu dul gis, te loolu dafay tax ab model def ay wax yu wóór te juum.
Résumé
Robustness is the field dedicated to defending against them, and it reveals deep gaps between machine and human perception.
Plongeur bu xóot
Ci 2013-2014, gëstukat yi dañu wane ni yokk benn motif bruit buñu defar bu baax, buñu daanaka kenn gisul ci benn nataal mën na soppi benn classifier ci 'panda' dem ci 'gibbon' ak wóolu bu rëy. Yii misaal yuy xeex dañuy jàpp ni reso neuronal yi dañuy jàng ay dogal yu yomba dagg ci espace bu am dimension yu kawe. Cong yi dañuy faral di nekk white-box (ki song xamna model bi te dafay jëfandikoo gradient yi, ni ci FGSM ak PGD) wala black-box (sorti yi kese lañuy gis). Lu yéeme mooy misaali noon yi dañuy faral di toxal ci diggante yeneen xeeti, loolu mooy tax ñu mëna song te kenn du ci dugg. Perte bi amul benn werante: ay sticker yu aduna physique mënnañu nax detektëru stop-sign yi, te 'jailbreak' yu gaaw yi ñooy analog modelu làkk wi. Gëstu Robustness dafay wut xeetu doxalin bu jaar yoon doonte ci jafe-jafe yu gëna bon, perturbations adversarial.
Gis-gis xarala
Cong yu bari dañu sukkandikoo ci gradient: FGSM dafay jël benn jéego ci wàllu màndarga gradient bu ñàkk ci wàllu duggal, fekk PGD dafay baamtu lii ci biir benn boule bu ndaw (lu melni, L-infinity) ci biir duggal bu njëkk bi. Kaaraange bi gëna am doole biñ xamee mooy tàggat noon, tàggataat ci misaali noon, ñu defaree ko ni jafe-jafe min-max: wàññi ñàkk ci perturbation bi gëna bon. Dafay gëna dëgër waaye dafay laaj njub ak xayma yu sell.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu misaali noon ak dëgër
Bi IA duggee ci sistem yu am solo ci wàllu kaaraange, dëgër mingi joge ci bëgg-bëggu akademik dem ci bëgg-bëggu ingenieur yi. Liggéey bi mungi weyal ci ay kaaraange yu sertifiye yuy garanti ci math ni benn perturbation ci biir benn yamale bu muna soppi genn gi, ak ci dëgër ci kaw njang mu gina yaatu, gina tar ci ay yamale yu jano ak ay xeetu lakk yu magg, lu mel ni jailbreak ak injection yu gaaw. Xaarandil benchmark adversarial buñ yamale, pipeline red-teaming, ak fitnaal buñ yamale ngir model yiñ dugal ci dawal seen bopp, kaaraange, ak pajum wérgi-yaram ngir wane wóor gu gëna bon.
Doxal ci àdduna dëgg
Gëstukat yi dañu def ay sticker yu ndaw ci kaw benn pano stop, loolu waral modelu vision bi juum ci li ko jàppee ni pano buy tënk gaawaay, loolu dafay wane ni oto yiy dawal seen bopp mën nañu am loraange ci àdduna bi.
Ekipu kaaraange yi dañuy xàmmee kanam gi ci ekipu xonk yi, ñu am ay patch yuñ imprime ci weer yi wala ci yéere yiy moytu wala ñuy nax nit ñi.
Spam ak filtre malware dañu leen di saytu ak ay dugal yuñ jaxasoo ngir moytu charges yu bonn yi fekk ñuy romb classifier yi.
Developpeur LLM yi dañuy xeex 'jailbreak' yu gaaw yi, làkk wi ñuy méngale ak misaali noon yi, di nax model yi ngir ñu baña topp tegtali kaaraange yi.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Adversarial Examples and Robustness quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Reseau adversaire yu defar
Laaj yi ñuy faral di laaj
What is Adversarial Examples and Robustness?
Misaalu noon yi ñooy ay dugal yuñ jaxasoo ci coppite yu ndaw yuñu dul gis, te loolu dafay tax ab model def ay wax yu wóór te juum. Robustness mooy wàll wiñ jagleel ngir di ko xeex, te dafay wane bërëb yu xóot yi am ci diggante masin ak gis-gis nit.
Luy misaalu noon?
misaali noon yi dañuy yokk ay perturbation yu ndaw, yuñ defar bu baax, yu nit ñi dul seetlu, waaye ñuy nax model bi ci njuumte yu wóolu seen bopp.
Lan mooy wuutale songu 'kees bu weex' ak songu 'kees bu ñuul'?
Ñiy song boyet bu weex xam nañu model bi te mën nañu jëfandikoo ay gradient; songe black-box yi duñu gis lenn ludul dugg ak genn.
Lan mooy kaaraange giñ gëna jëfandikoo ngir gëna dëgëral noon yi?
Taggat yaram dafay yokk jàng ak ay dugal yu gëna bon, ñu defaree ko ni optimisation min-max, te mooy kaaraange gi gëna am doole te wóor, waaye mën na wàññi njubte bu sell.
Lan moo waral 'mën toxal' misaali noon yi?
Ndax misaali noon yi dañuy toxal ci model yi, songekat bi mën na leen defar ci modelu surrogate ba noppi song ci anam wu jaar yoon benn target buñu mënul saytu.
Lan mooy misaalu làkk wu yaatu ci misaali noon?
Jailbreaks ak injections yu gaaw yi ay dugal yuñ defar lañu yuy jaay xel LLM ci jeffin ju wóorul wala juñu bëggul, di méngoo ak songe noon yi ci gis-gis.