Chii chaitika
Vatsvaguri vanotsanangura iyo Groundhog Bit-Flip Attack, iyo inonangana nzira dzekufambisa dzinosarudza kuti ndeipi nyanzvi sub-network musanganiswa-we-nyanzvi mutauro modhi inoita. Iro bepa rinoti kupenengura nhamba diki yenzira-yakapetwa mabheti kunogona kuita kuti mamodheru aburitse zvakati rebei zvinobuda asi zvakanyanya kuchengetedza zvazvinoreva.
Gwaro rearXiv, rakatumirwa muna Nyamavhuvhu 26, 2026, rinopa zvinodaidzwa nevanyori varo kuti Groundhog Bit-Flip Attack, kana GBFA. Bepa rinotarisa pamusanganiswa-we-nyanzvi, kana MoE, mhando dzemitauro. Mune aya masisitimu, nzira yekufambisa inosarudza inomutsa nyanzvi sub-network kune akasiyana tokeni. Vanyori vanopokana kuti chimiro ichi chinogadzirisa chinogadzira nzvimbo nyowani yekurwisa nekuti dzimwe nyanzvi dzinogona kusarongeka dzakabatana nemamwe ma tokens, kusanganisira ekupedzisira-yekutevedzana tokeni. Chirevo chepakati ndechekuti munhu anorwisa anogona kushandisa masangano aya nekushandisa mabheti muchikamu chenzira.
Iro bepa rinoratidzira GBFA sediki-flip-yakavakirwa "Denial-of-Wallet kuwanikwa kurwisa" kurwisa MoE-based LLMs. Mune mazwi anoshanda anotsanangurwa neabstract, kurwiswa kwacho kunoitirwa kuita kuti decoding ienderere mberi kwenguva yakareba kupfuura yakajairwa, kuwedzera mashandisirwo etokeni uye kusundidzira zvakawanda zvinobuda kune yakanyanya system tokeni muganho. Kwainobva hakutsanangure zita iri mune zvemari kunze kwekubatana kwaro kune yakawedzerwa decoding, uye haripe yakayerwa mutengo wedhora, sevhisi-level maitiro, kana fungidziro yekuti kangani kakaiswa hardware inogona kuwana inodiwa bit flips.
Sekureva kwebepa, vaongorori vakaedza nzira iyi kune ina "main real-world" MoE-yakavakirwa mitauro yemhando uye nepamusoro pekutaura, kufunga, uye ejenti mabasa. Vanoshuma kuti nemaoko kudzima avhareji yevasingasviki ina nyanzvi kwakagadzira avhareji yekukwira kwemitengo ye5,912%, nehuwandu hwemasampula ebvunzo anosvika pahupamhi hwechiratidzo. Iyo abstract inotaurawo kuti semantic kutendeka kwakachengetedzwa zvakanyanya, zvichireva kuti zvakabuda zvakaramba zvine chekuita nebasa rakakumbirwa pane kungove zvisizvo. Izvi zvirevo kubva muongororo yakashumwa yebepa; iyo yakapihwa hairatidze mamodheru, inopa saizi dzemuenzaniso, rondedzero yekutanga uye yakarwiswa tokeni kuverenga, kana kutsanangura kuyedza hardware uye kukanganisa-jekiseni kuseta.
Nei zvichikosha
Mhedzisiro yakashumwa inoratidza njodzi-yakasarudzika yekuwanikwa inosiyana nekurwiswa kwakanangana nekushandura zvinotaurwa neAI system. Kana yakasimbiswa kunze kwekuyedzwa kwebepa, diki hardware kana chikanganiso chekuyeuka chinogona kukonzera kuremerwa kwemabasa kuti adye yakawandisa decoding kugona kupfuura zvinotarisirwa. Kunobva kwacho hakuratidze kuti kurwiswa kunoshanda sei kune masisitimu akaiswa kana kuyera mutengo wayo wemari.
Hurukuro zhinji dzinozivikanwa dzekurwiswa kwemhando yemutauro dziri pakuchinja zvirimo zvemodhi, kudarika chengetedzo, kutora ruzivo, kana kushandura mushandisi. Iri bepa rinotsanangura imwe nzira yekutadza: sisitimu inogona kuenderera mberi ichiburitsa zvakabatana zvinobuda ichiri kuita zvishoma. Musiyano iwoyo unokosha nekuti wakajairika mhando cheki inogona kupotsa dambudziko. Mhinduro inogamuchirwa semantic inogona kuramba ichiisa basa rakakura risingatarisirwe rekudhikodha.
Huyero yakashumwa yemhedzisiro yakakosha kana inobata mumaseti ekushanda. Avhareji ye5,912% yeavhareji yekuwedzera kwehurefu hwekubuda ingareva kushandiswa kwakanyanya kweiyo inference kugona kune zvakakanganisika zvikumbiro, nepo zvinobuda zvinopinda mupimo wezviratidzo zvinogona kutora vashandi kana mitsetse kwenguva yakareba. Mhedzisiro yacho inonyanya kukosha kune masisitimu anoshandira vashandisi vazhinji, anomhanya akazvimiririra kana agentic workflows, kana bhajeti compute zvichienderana nehurefu hwekupindura hunotarisirwa. Nekudaro, sosi haaratidze kuti kurwiswa kwacho kwaitwa pakurwisa sevhisi yekugadzira, kuti inogona kukonzeresa kure, kana kuti yaizokonzeresa kubuda.
Zvakawanikwa zvakare zvinonongedza kune dhizaini yekutengeserana muMoE masisitimu. Sarudzo yehunyanzvi activation inoshandiswa kuyera mamodheru emitauro zvakanaka, asi chimiro chimwechete chenzira chinogona kugadzira kutsamira kwakasimba: boka diki renyanzvi rinogona kuve nemhedzisiro yekumisa kana humwe hunhu hwechiratidzo. Kana dudziro yepepa iri chokwadi, kuyedzwa kwekusimba kwemamodhi eMoE kungangoda kuongorora kwete chete chokwadi uye zvinokuvadza zvinobuda asiwo zvisina kujairika mashandisirwo etokeni uye kutadza kwenzira. Sosi inotsigira izvi sechirevo chekuchengetedza, kwete sehumbowo hwekuti zvese zvivakwa zveMoE zvinogovana kushaya simba kwakafanana.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibvunzo yakakosha inotevera ndeyekuti mhedzisiro yacho inoberekana kune mamwe emamwe musanganiswa-we-nyanzvi modhi, ndeapi masvikiro kana mamiriro ekukanganisa anodiwa kupepeta mabhiti akakodzera, uye kuti kuchengetedzwa kwakajairwa kunoona here kana kudzikamisa maitiro. Sosi haritauri mamodheru mana akaongororwa, kutsanangura kudzikisira, kana kupa humbowo kubva mukuyedzwa kwakazvimirira.
Nyaya yekutanga kutarisa ndeye reproducibility. Iyo arXiv peji inoti kuyedza kwakavhara mana maMoE-based LLMs, asi rakapihwa sosi hariadoma mazita kana kutsanangura mavhezheni avo. Kumwe kuongorora kunofanirwa kuona kana mhedzisiro yacho inoenderana neamwe magadzirirwo enzira, nzira dzehuwandu, marongero endangariro, kuverenga kwenyanzvi, kana hunhu hwekumisa. Mibairo pamamwe masisitimu akavhurika uye akaisirwa zvaizobatsira kusiyanisa kushaya simba kwekuvaka kubva pakusagadzikana kunogumira kune mamwe maitirwo.
Nyaya yechipiri ndeyekurwisa kugona. Iyo abstract inoreva kuziva uye kupenengura routing-layer bits uye kune mawoko kudzima nyanzvi, asi haitaure kuti ndeipi nhanho yekuwana inodiwa neanorwisa, mabits anosvika sei, shanduko dzichienderera mberi, kana kuti zvakajairwa zvibvumirano zvesoftware zvinogona kuvadzivirira. Izvo zvakare hazvitaure kana iyo yekutyisidzira modhi inosanganisira yenguva pfupi yekukanganisa kwehardware, kugadziridzwa kwakashata kwememodhi yekurangarira, zvivakwa zvakakanganisika, kana imwe nzira. Iwo mameseji anoona kuti GBFA inonyanya kutsvaga kusimba kwerabhoritari kana kutyisidzira kunoitika nekukurumidza.
Pakupedzisira, vashandi nevatsvakurudzi vachada humbowo pamusoro pekudzivirira. Iyo sosi haitauri kudzikisira, yekuona chikumbaridzo, maitiro ekudzoreredza, kana kuenzanisa neyakajairwa mareti emuganho uye yakanyanya-kubuda zvinodzora. Basa rinobatsira rekutevera raizoyedza kutarisa kune isingawanzo ratidziro inflation, nyanzvi yekudzima, uye yakadzokororwa kureba-kupedzisa kupedzisa, uku kuyera maaramu enhema uye mutengo wekuita. Kudzokorora kwakazvimiririra, kujekesa kwakajeka kwemamodeli akaongororwa uye zvigadziriso, uye kuyedzwa pasi pemamiriro ekushandira chaiwo kwaizodiwa usati washandura mhedzisiro yepepa yakashumwa muhuwandu hwehuwandu hwengozi.