- I-GPUHammer ikhombisa ukuthi kungenzeka ukwehlisa ukunemba kwamamodeli e-AI kube amazinga angenamsebenzi ngokusebenzisa ubuthakathaka obungokomzimba kwimemori ye-GPU.
- Ukuphepha kwe-AI kuhlanganisa ihadiwe, idatha, amamodeli, kanye nengqalasizinda, futhi kuhlanganisa ukuzivikela ekuhlaselweni okuphikisayo kanye nobunjiniyela bezenhlalo obusebenzisa i-AI.
- Imithwalo yemisebenzi ye-AI idinga amasu athile okuvikela, okuhlanganisa ukusetshenziswa kwe-ECC, ukuhlukaniswa kwe-GPU, ukuqapha okuqhubekayo, kanye nokulawula umjikelezo wokuphila wemodeli.
- I-AI iyinkemba esika nhlangothi zombili ekuphepheni kwe-inthanethi: ithuthukisa kokubili amakhono okuzivikela nawokuhlasela, ngakho-ke inzuzo isekusu nasekubuseni.
Ukuphepha kobuhlakani bokwenziwa akusekho isihloko sezemfundo kuphela futhi sekuyinkinga yangempela ezinkampanini, abacwaningi, kanye nohulumeni bomphakathi. Asikhulumi nje nge-malware noma i-phishing yendabuko: manje sekukhona ukuhlaselwa okungakwazi ukwehlisa ngokuphelele ukunemba kwemodeli ye-AI , ukusebenzisa ubuthakathaka behadiwe ebonakalayo, ngisho nokulawula umbono womuntu ngokusebenzisa ama-deepfake kanye nemikhankaso yolwazi oluzenzakalelayo.
Ngaphakathi kwalesi simo esisha, i-GPUHammer ivelele, ukuhlasela okunciphisa ukunemba kobuhlakani bokwenziwa kube amazinga aphansi kakhulu ngokusebenzisa ubuthakathaka kwimemori ye-GPU. Kodwa akulona kuphela usongo: kukhona nokuhlaselwa kwezitha kwedatha, ubunjiniyela bezenhlalo obusebenzisa i-AI, ukucwasa kwe-algorithmic okunomthelela kwezenhlalo, kanye nengqalasizinda ye-AI, ezimweni eziningi, engaklanywanga ngokuphepha engqondweni. Ake sikuchaze ngokuthula, kodwa ngokujula kobuchwepheshe okudingayo.
Kuyini i-GPUHammer futhi kungani ingabhubhisa ukunemba kwemodeli ye-AI?
I-GPUHammer uhlobo lokuhlasela kwe-Rowhammer , inqubo esebenzisa iphutha elibonakalayo kumaseli enkumbulo. Ngenkathi ukuhlaselwa kwe-Rowhammer okuvamile kugxile ku-DRAM esetshenziswa ama-CPU, i-GPUHammer idlulisela ukugxila kwimemori ye-GPU , okuyiyona nto eyinhloko yokuqeqeshwa kanye nokucabanga kwamamodeli obuhlakani bokwenziwa besimanje.
Ukuhlasela kusebenza ngokuphindaphindiwe nangendlela elawulwayo ukufinyelela imigqa ethile yememori ye-GPU. Le misebenzi eqinile nevumelaniswe ngokucophelela ibangela ukuvuza kweshaja kagesi okubangela ukuthi amabhithi aseduze ashintshe inani. I-0 iba yi-1 noma ngokuphambene nalokho, ngaphandle kokuthi uhlelo luyibone njengokwehluleka okusobala, ngoba ibonakala njengamaphutha athule kudatha.
Uma kukhulunywa ngemodeli ye-AI, lawa ma-bits ashintshiwe angathinta amapharamitha emodeli kanye nokusebenza okuphakathi noma idatha yokufaka ecutshungulwa yi-GPU. Ngokonakalisa izisindo noma ama-tensor ayisihluthulelo, imisebenzi yezibalo ayithembeki futhi izibikezelo ziyehla ngekhwalithi. Okukhathazayo kakhulu ukuthi uhlelo luyaqhubeka nokusebenza ngendlela evamile, kodwa ukunemba kwemodeli kwehla kakhulu.
Abacwaningi ababonise i-GPUHammer bachaza ukuthi i-attack vector ingasebenzisa izindawo zokubala ezifana ne-CUDA ukufaka ukufinyelela kwememori ethile nokulawula iphethini yokushaqeka. Ngokusebenzisa amandla aphezulu okusebenza afanayo ama-GPU anamuhla, inani lokufinyelela lingakhuliswa ngesikhathi esifushane kakhulu, okwandisa amathuba okuthola ama-bit flips awusizo kumhlaseli.
Inselele enkulu kulabo abakha lokhu kuhlasela kube ukungabonakali kokusetshenziswa kokuphepha kwimemori ye-GDDR , ukumepha kwekheli langaphakathi, kanye nokubambezeleka okuhlukile uma kuqhathaniswa ne-CPU DRAM. Ngokwenza lokhu kumepha kabusha, ukuvumelanisa nama-windows okuvuselela imemori, kanye nokwenza ngcono ukufana, bakwazile ukunqoba lezi zithiyo zethiyori.
Umthelela we-GPUHammer: kusukela ekunembileni okungu-80% kuya ku-0,1%
Icala elihlaba umxhwele kakhulu elibhalwe ngabacwaningi yilelo le- NVIDIA A6000 , i-accelerator esezingeni eliphezulu evame ukusetshenziswa emisebenzini yobungcweti ye-AI. Imodeli evame ukuthola ukunemba okungu-80% emsebenzini wayo yehliswe yaba ngu -0,1% omncane ngemuva kokusebenzisa ukuhlaselwa kwe-GPUHammer.
Sikhuluma ngokwehla kokusebenza okwenza i-GPU ibize ama-euro ayizinkulungwane eziningana ngaphezu kwesisindo sephepha sokunethezeka uma kukhulunywa ngemisebenzi ye-AI. Imodeli enokunemba okungu-0,1%, empeleni, ayisebenzi kunoma yiluphi uhlelo lokusebenza olubalulekile: ukuxilongwa kwezokwelapha, izinhlelo zokuthola ukukhwabanisa, ukubikezela kwezezimali, noma ukuhlukaniswa kwezithombe kushintsha kusuka ekubeni amathuluzi okusekela izinqumo kuya ekubeni abakhiqizi bomsindo nje.
Ezindaweni zebhizinisi, lokhu kusho ukuthi amamodeli angenza izinqumo ezingalungile kakhulu ngaphandle kokuthi isizathu sibonakale ngokushesha. Uhlelo lokuthola amaphuzu esikweletu lungavumela abafakizicelo abangafaneleki noma lwenqabe amakhasimende afanele isikweletu; uhlelo lokuphepha kwe-inthanethi olusekelwe ku-AI lungase luphuthelwe izinsongo zangempela futhi ludale imiphumela engamanga yonke indawo. Umthelela uholela ekulahlekelweni kwezezimali, ukulimala kwedumela , ngisho nemiphumela yezomthetho engaba khona uma kubonakaliswa ukunganaki ekuvikelweni kwengqalasizinda.
Ngokuthakazelisayo, ukuhlaselwa akuhambisani nawo wonke amakhadi. Ukuhlolwa kubonisa impumelelo kuma-GPU afana ne- NVIDIA A6000 ene-GDDR6 , kodwa hhayi kumamodeli afana ne-RTX 3080. Inkolelo-mbono iwukuthi, ngokusebenzisa imemori evela kubakhiqizi abahlukene, ngezindlela ezahlukene zokunciphisa kanye namamephu , ukusebenza kokuhlaselwa kuyahlukahluka. Lokhu kusikisela ukuthi imiklamo ethile yememori ingase ibe namandla kakhulu ngokwemvelo.
Ukuze bahlole futhi babhale phansi ukuhlaselwa, abacwaningi bashicilele imininingwane yobuchwepheshe kwiwebhusayithi esemthethweni yale phrojekthi, lapho bechaza khona inqubo, imikhawulo yamanje, kanye nokwandiswa okungenzeka esikhathini esizayo, begcizelela ukuthi lolu hlobo lokuhlaselwa kwama-GPU lumelela umngcele omusha ekuphepheni kwehadiwe eqondiswe ekuhlakanipheni kokwenziwa.
Ukuvela kokuhlaselwa kweRowhammer: kusukela ku-ZenHammer kuya ku-GPUHammer
I-GPUHammer ayizange ivele ngokuzumayo; iyingxenye yomugqa wokuziphendukela kwemvelo wokuhlaselwa kweRowhammer obekukhathaza umphakathi wezokuphepha iminyaka eminingi. Ngo-2024, kwembulwa iZenHammer, ukuhlasela okuklanyelwe ukuxhaphaza ubuthakathaka kuma-CPU e-AMD Ryzen kanye ne-EPYC, okubonisa ukuthi isakhiwo sangaphakathi se-DRAM sivumele ukuthi ama-bit flips aqalwe ngokushaya imigqa eseduze.
Ucwaningo olulandelayo luveze izinkinga ezifanayo kwimemori yakudala ye-DRAM kanye nasezizukulwaneni zakamuva njenge-DDR5 ezivela kubakhiqizi abathile. Unyaka nonyaka, kuye kwatholakala ama-vector amasha noma ukuhlukahluka komgomo ofanayo: ukuphoqa ngokomzimba imemori ukuthi iziphathe kabi ngaphansi kwamaphethini okufinyelela athile kakhulu.
Umehluko oyinhloko ukuthi, kuze kube manje, izifundo eziningi zigxile kakhulu kwimemori eyinhloko ye-CPU. Nge-GPUHammer, ukugxila kushintshela kakhulu kwimemori yamakhadi ehluzo , isici esibalulekile sokucubungula okukhulu okuhambisanayo okudingekayo ekufundeni okujulile. Lokhu kusho ukuthi yonke ingqalasizinda esekelwe ku-GPU ye-AI—kusukela kumaseva esikhungo sedatha kuya ezindaweni zokusebenza kanye namakhadi ehluzo abathengi abasezingeni eliphezulu—manje isengozini yalezi zinsongo.
Lokhu kuqhubeka kwenza kucace ukuthi umncintiswano phakathi kwabakhiqizi behadiwe nabacwaningi bezokuphepha ngeke uphele. Isizukulwane ngasinye esisha senkumbulo (GDDR6, GDDR7, HBM3, njll.) senezela amakhono okusebenza, kodwa futhi nezindawo ezintsha zokuhlasela okumele zihlolwe kusukela esigabeni sokuklama, hhayi nje kuphela ngemva kweqiniso.
Ukunciphisa kwamanje: I-ECC, izinkumbulo ezintsha kanye nezindleko zazo
Umzila wokuqala wokuzivikela ekuhlaselweni kohlobo lwe-GPUHammer ukuvumela inkumbulo yokulungisa amaphutha (ECC) noma nini lapho itholakala. I-ECC ingeza ama-bit engeziwe kudatha ukuthola futhi, ezimweni eziningi, ukulungisa amaphutha e-single-bit noma ama-multi-bit ngesikhathi sangempela, inciphise umphumela wokuphenduka okuningi okubangelwa ukuhlaselwa.
Inkinga iwukuthi ukuvumela i-ECC akuveli ngaphandle kwentengo: izifundo zibonisa ukulahlekelwa kokusebenza okulinganiselwa ku-10% kwamanye ama-GPU uma lesi sici sivuliwe. Ezinhlanganweni ezimatasa ngokucindezela yonke ithonsi lokugcina lokusebenza kuyo yonke i-teraFLOP, lesi sijeziso singase singabi lula, kodwa yintengo okufanele ikhokhwe ukuze kuthuthukiswe kakhulu ubuqotho bedatha.
Ezizukulwaneni ezintsha zememori njenge -GDDR7 ne-HBM3 , abakhiqizi abaningi sebevele behlanganisa izindlela ze-ECC ngqo kuma-memory chips, benciphisa ubungozi bokuhlaselwa yiRowhammer kanye nezinhlobo zazo. Lobu buchwepheshe buklanywe ngokuqonda ukuthi imisebenzi ye-AI ayidingi nje isivinini kodwa futhi nokuqina ngokumelene namaphutha athule , kungakhathaliseki ukuthi avela ekuhlulekeni ngokomzimba noma ekuhlaselweni okunonya.
Abacwaningi abethule i-GPUHammer babonisa ukuthi bayaqhubeka nokusebenza kumasu amasha okuzama ukugwema ngisho nalezi zivikelo , befuna izindlela ezihlukile lapho ukulungiswa kwamaphutha kunganele noma lapho kungenzeka khona ukuphoqa izimo ezimbi kakhulu ezinqoba izindlela ze-ECC.
Ngokwabo, abakhiqizi abanjengoNVIDIA baziswa ngobuthakathaka kusenesikhathi (kulokhu, cishe ngoJanuwari), futhi batusa ngokusobala ukuthi i-ECC ivunyelwe ezindaweni ezibucayi, kanye nokusebenzisa imikhuba emihle yokuhlukaniswa komsebenzi lapho kwabelwana ngezinsiza ze-GPU phakathi kwamakhasimende noma izinhlelo zokusebenza.
Ukuphepha kwe-AI: ngale kwehadiwe kanye ne-GPUHammer
Nakuba i-GPUHammer iheha izihloko eziningi, ukuphepha kwe-AI kuhlanganisa ububanzi obubanzi kakhulu kunokuhlaselwa kwememori ebonakalayo. Imithwalo yemisebenzi ye-AI yethula izindawo zokuhlasela amasu endabuko okuphepha kwe-cyber angazange azihlanganise ngokugcwele: idatha yokuqeqesha, amamodeli, amapayipi okucabanga, ama-API esevisi, kanye ne-GPU kanye nengqalasizinda ye-accelerator uqobo.
Ukuphepha kwe-AI kugxile ekuvikeleni le mithwalo yemisebenzi ekusetshenzisweni noma ekuphathweni kabi, ukuqinisekisa ukuthi amamodeli asebenza njengoba kuhlosiwe futhi ngaphandle kokuphazamiseka kwangaphandle. Lokhu kuhilela ukuvikela ubuqotho bedatha yokuqeqesha, ukuqinisekisa ukuvela kwemodeli (ukwazi ukuthi ubani owayiqeqesha, kanjani, nokuthi yini), kanye nokusungula izindlela zokuhlukanisa ama-GPU nezinye izisheshisi ezisekelweni lapho ziphethwe khona.
Amasu amaningi okuphepha kwezinkampani aklanyelwe ukuvikela izinhlelo zokusebenza zendabuko: amaseva ewebhu, izizindalwazi, izindawo zokusebenza. Ngokuphambene nalokho, ikhambi le-AI lifana kakhulu "nesakhiwo sefulethi kwenye iplanethi" : amamodeli amaningi, imithombo yedatha ehlukahlukene, ama-API angaphandle, amapayipi e-MLOps, ama-GPU abiwe efwini... kanye nenani elikhulu lamaphuzu okungena okungenzeka ukuthi awazange abhekwe njengezivikeli zokuhlasela.
Ukuze i-AI ithembeke ezingeni lebhizinisi, kudingeka isu elithile lokuphepha elihambisana nezilawuli ezikhona: ukuhlukaniswa kwenethiwekhi, ukuphathwa kobunikazi, ukuthobela imithetho, ukuqapha okuqhubekayo, kanye nokuphendula kwezehlakalo. Ukwengeza nje olunye uhlelo lokuvikela akwanele; kubalulekile ukuqonda ukuthi idatha igeleza kanjani isuka emthonjeni wayo iye ekucabangeni, nokuthi ingashintshwa kuphi.
Ingqalasizinda ye-AI: ubuthakathaka obusha kanye nezidingo zokuvikelwa
Uma sicabanga ngengqalasizinda ye-AI, asikhulumi nje ngamaseva anamandla: sikhuluma nge -ecosystem ephelele ehlanganisa ama-GPU, isitoreji esisebenza kahle kakhulu, amanethiwekhi okubambezeleka okuphansi, amapulatifomu okuqondisa izitsha, izinsizakalo zamafu, namathuluzi e-MLOps. Isendlalelo ngasinye sengeza ukusebenza, kodwa futhi namathuba amasha abahlaseli.
Enye yezinkinga ezibalulekile ukuntuleka kokuhlukaniswa okuqinile kuma-GPU amaningi . Ngokomlando, amakhadi ehluzo awaklanywanga ngezinga elifanayo lokuhlukanisa kanye nezindlela zokuphepha njenge-CPU, okwenza abe yizinhloso ezikhangayo. Ukuhlaselwa okufana ne-GPUHammer kusebenzisa leli qiniso: amalogi amabi, ukubonakala okulinganiselwe ekufinyeleleni imemori, kanye nendawo eyimfihlo engabonakala kalula.
Ngaphezu kwalokho, ezinhlanganweni eziningi, izinsiza ze-GPU zabelwana phakathi kwamaqembu ahlukene, amaphrojekthi, noma ngisho namakhasimende (ezindaweni ezinabantu abaningi abaqashayo). Ngaphandle kokuhlukaniswa okucacile kanye nokulawula okuqinile kokufinyelela , umthwalo womsebenzi ononya ungathinta amamodeli anenani eliphezulu asebenza emshinini ofanayo ongokoqobo.
Isu lokuphepha le-AI elihlelwe kahle aligcini nje ngokuvikela idatha ebucayi ekuphumeni, kodwa futhi lisebenzisa izinyathelo zokuvimbela ukuhlaselwa kwamamodeli, liqinisa ukuthobela imithetho (isb., uMthetho we-EU AI), futhi lihlanganisa amakhono okuqapha ukuthola okungahambi kahle ekusetshenzisweni kwemithwalo yemisebenzi ye-AI.
Ekugcineni, uma uqonda kangcono ubuchwepheshe obuyisisekelo—indlela i-AI eqeqeshwa ngayo, esetshenziswa ngayo, futhi ekhuliswa ngayo—kulapho kuba lula khona ukubona nokuvala lezo “zicabha namafasitela” ezengeziwe ezingekho engqalasizinda ye-IT yendabuko.
Ukuhlaselwa kwabaphikisi: ukuphatha idatha ukuze kudideke i-AI
Ngaphandle kwehadiwe, i-AI isengozini enkulu yokuhlaselwa okuphambene , okugxila ekuphatheni okokufaka kwemodeli (izithombe, umsindo, umbhalo, idatha yethebula) ngokusebenzisa ukuphazamiseka okuncane kakhulu kodwa okuklanywe ngokucophelela ukuze kubangele amaphutha okuhlukaniswa noma izinqumo ezingalungile.
Empeleni, umhlaseli angafaka izinguquko ezingabonakali emehlweni omuntu esithombeni se-panda futhi abangele ukuthi imodeli iyibize ngokuthi isilwane esihluke ngokuphelele. Kungenzeka okufanayo ngezimpawu zomgwaqo ezimotweni ezizimele , izithombe zezokwelapha ekuxilongweni okusiza ngekhompyutha, noma imibhalo esetshenziswa izinhlelo zokuthola ukukhwabanisa.
Lokhu kuhlasela kusebenzisa iqiniso lokuthi amamodeli okufunda okujulile asekela ukusebenza kwawo kumaphethini ezibalo aphezulu, hhayi "ekuqondeni" umongo njengomuntu. Amasu afana ne- Fast Gradient Sign Method (FGSM) , i-Projected Gradient Descent (PGD), noma ukuhlasela kwe-Carlini & Wagner kubala ukuphazamiseka okuthi, ngokushintsha okuncane kumaphikseli noma amavekhtha okufaka, kwandisa amathuba okwenza imodeli iphambe.
Izingozi zisobala: endaweni yokunakekelwa kwempilo, ukuphathwa kabi kwe-MRI scan kungaholela ekuxilongweni okungalungile ; ekuphepheni kwe-inthanethi, izibonelo eziklanywe ngokukhethekile zingavumela i-malware ukuthi ingatholakali yizihloli ezisekelwe ku-AI; ekubonweni kobuso, izinguquko ezincane esithombeni zingazenza sengathi zingobani noma zigweme izilawuli.
Eqinisweni, imiboniso yangempela isivele ikhona: ukuhlaselwa kwabahlukanisi bezithombe abavela kubahlinzeki abakhulu bezobuchwepheshe, imiyalo yezwi efihliwe kumsindo obonakala ungenangozi wabasizi bedijithali, noma ama-imeyili ashintshwe kancane ukuze kugwenywe izihlungi zogaxekile ngokusekelwe ekufundeni komshini.
Ukuzivikela ekuhlaselweni yizitha: ukuqina nokutholwa
Ukuvikela i-AI ekuhlaselweni okuphambene kudinga indlela enezigaba eziningi. Enye yamaqhinga afundwe kakhulu ukuqeqeshwa okuphambene , okuhlanganisa ukufaka izibonelo ezimbi ngesikhathi sokuqeqeshwa kwamamodeli ukuze amamodeli afunde ukuqaphela nokumelana nalezi zinhlobo zokuphazamiseka. Nakuba lokhu kuthuthukisa kakhulu ukuqina, kwandisa nezindleko zokubala futhi akuqinisekisi ukumbozwa kwazo zonke izithasiselo zokuhlasela ezingacatshangwa.
Esinye isivikelo ukucubungula okufakwayo kanye nokuqinisekisa . Imisebenzi efana nokuhlunga, ukushelela isithombe, kanye nokunciphisa umsindo kunganciphisa ukuphazamiseka, kodwa kumele kusetshenziswe ngokucophelela ukuze kugwenywe ukwehlisa ukunemba kwemodeli ezimweni ezisemthethweni. Ngaphezu kwalokho, amasu okuthola okungavamile kanye nokuhlaziywa kwezibalo kusetshenziswa ukuhlonza amaphethini okufaka asolisayo.
Ucwaningo luyenziwa futhi ezakhiweni eziqinile kakhulu, njengezindlela zokuzivikela noma izindlela zokuhlanganisa (ukuhlanganisa amamodeli amaningana nokuqhathanisa imiphumela yawo), okwenza kube nzima ukubikezela ngqo ukuthi uhlelo luzosabela kanjani ezinkingeni ezincane. Umgomo ukunciphisa ukuzwela kwemodeli kulezi zinguquko ezincane kakhulu.
Noma kunjalo, ukuzivikela ekuhlaselweni yizitha kuwumdlalo wekati negundane: yonke intuthuko yokuzivikela ivame ukulandelwa amasu amasha okuhlasela azama ukuyigwema. Ngakho-ke, ngaphezu kokuthuthukisa amamodeli, kubalulekile ukusungula izilawuli zomongo: ubani ongathumela uhlobo luni lwedatha, kusiphi isiteshi, ngaphansi kobuphi ubuqiniso, nokuthi imiphumela engavamile iqashwa kanjani.
Uma sibheka esikhathini esizayo, i-AI (XAI) echazekayo ithola ukubaluleka, okusiza ukuqonda ukuthi kungani imodeli yenza izinqumo ezithile futhi ithole amaphethini okuziphatha angajwayelekile angahlotshaniswa nokuhlasela kwesitha okungaziwa okwamanje.
Akukhona konke mayelana namamodeli amancane kanye nezibalo; ingxenye ebalulekile yosongo ivela kubunjiniyela bezenhlalo obuthuthukisiwe yi-AI . Abahlaseli basebenzisa amamodeli olimi, izinhlelo zokuhlaziya idatha, kanye namathuluzi okukhiqiza okuqukethwe ukuze baklame ukukhwabanisa okwenziwe ngezifiso kakhulu, okukholekayo, futhi okunzima ukukubona.
Lokho okwakudinga isikhathi, ikhono lengqondo, nokubhala okuhle—ukucwaninga isisulu, ukuqonda umongo waso, nokubhala umlayezo othakazelisayo—manje sekungenziwa ngokuzenzakalela. I-AI yenza kube nokwenzeka ukuqoqa ulwazi lomphakathi ezinkundleni zokuxhumana, kumawebhusayithi ezinkampani, kanye nokuvuza, ukuphrofayela okuhlosiwe, nokukhiqiza ama-imeyili, imiyalezo, kanye nezikripthi zokushaya ucingo ngemizuzwana ngezwi eliqondiswe ezweni ngalinye, emkhakheni, noma emasikweni.
Ochwepheshe bezokuphepha kwe-inthanethi bakhomba ukuthi izithiyo zolimi namasiko seziphelile cishe kubahlaseli. Isigebengu singacela umsizi wengxoxo osebenzisa i-AI ukuthi abhale i-imeyili ethembekile kubasebenzi benkampani eSpain, egcwele izinkomba ezinembile zemboni kanye nethoni yenkampani ekholekayo, ngaphandle kwesidingo sokukhuluma kahle ulimi.
Abacwaningi be-Threat bakhomba ukuthi leli khono selivele lisetshenziswa emikhankasweni emikhulu yobugebengu bokweba imininingwane ebucayi, ukukhwabanisa, kanye nokusatshalaliswa kwe-malware , lapho intuthuko ebaluleke kakhulu isezingeni lokubhala kanye nokungabikho kwamafulegi abomvu ajwayelekile (amaphutha esipelingi, izinkulumo ezingavamile, njll.). Amakhasi obugebengu bokweba imininingwane ebucayi nawo ayazuza emiklamo ephucuzekile, enezinkomba ezimbalwa ezibonakalayo zokukhwabanisa.
Kulesi simo, ukuhlaselwa akusabi yizithombe ezingahleliwe futhi kuba imikhankaso ehlanjululwe kakhulu , enezinga lokwenza kube ngokwakho elenza umamukeli abone umlayezo njengengxenye yemvelo yenqubo yakhe yedijithali.
Ngaphakathi kwalolu hlelo lokusebenza, kunezinhlobo eziningana zokuhlaselwa eziye zandiswa ubuhlakani bokwenziwa:
Ngakolunye uhlangothi, kukhona ubugebengu bokweba imininingwane obuqondene nomuntu siqu kusetshenziswa amamodeli olimi , lapho i-AI ihlaziya khona idatha yomphakathi futhi ikhiqize imiyalezo ehambisanayo, ebhalwe kahle ehambisana nendima kanye nomongo wesisulu. Amaphutha ayisisekelo ayanyamalala, futhi ama-imeyili abonakala njengokuxhumana kwangaphakathi okusemthethweni mayelana nokubuyekezwa kokuphepha, izinqubo ze-HR, noma izinkinga zobuchwepheshe.
Okuqondiswe kakhulu yi- automated spear phishing , egxile kubantu abathile noma amaqembu ngaphakathi kwenhlangano (ukuphathwa kwezezimali, abaphathi bezinhlelo, abaphathi abaphakathi). Izinkomba ze-AI zixhumanisa amaphrofayili omphakathi, amazinga ezinkampani, kanye nobudlelwano bokwethembana ukudala imiyalezo ehambisana namaphrojekthi angempela kanye nokuguquguquka kwangaphakathi okwaziwayo.
Emkhakheni wobuchwepheshe bezwi, i-vishing nge-cloning yezwi ithatha igxathu elikhulu phambili. Amamodeli e-AI angaphinda i-timbre kanye ne-pination yomuntu wangempela esebenzisa imizuzwana embalwa yomsindo. Ucingo olubonakala sengathi luvela ku-CEO, umhlinzeki oyinhloko, noma isikhulu sezokuphepha lungacindezela isisulu ukuba senze ukudluliselwa noma ukushintsha iziqinisekiso ngaphandle kokuqinisekisa ngezinye iziteshi.
Enye indlela yokwenza izinto iwukushushumbisa ngobuhlakani (ubugebengu bokweba imininingwane nge-SMS noma imiyalezo esheshayo), lapho i-AI ihlaziya khona imikhuba yokusebenzisa ifoni bese ikhetha isikhathi, ithoni, kanye nokuqukethwe kwemiyalezo okungcono kakhulu ukuze ikhohlise isisulu ukuthi sichofoze isixhumanisi, silande uhlelo lokusebenza olunobungozi, noma sinikeze iziqinisekiso. Imiyalezo ifihlwa njengezaziso zokulethwa kwephakheji, izaziso zasebhange, noma ukuxhumana okuvela kumasevisi avamile.
Okokugcina, ukusetshenziswa kwama -deepfakes ekuzibonakaliseni njengomphathi kwandisa ingozi ezindaweni lapho kunezingcingo zevidiyo ezivamile. Ividiyo ebonakala ingokoqobo yomphathi ecela isinyathelo esiphuthumayo idala umbono obonakalayo nozwakalayo wobuqiniso onqoba izithiyo eziningi zokuqinisekisa uma izinqubo ezicacile zingekho.
Ama-chatbot anonya, ukwebiwa kobunikazi, kanye nolwazi olungelona iqiniso
I-AI yengxoxo isibe yisikhali sabahlaseli. Ama-chatbot anonya anokucubungula ulimi lwemvelo angazenza izinsizakalo zokusekela, abasizi bangaphakathi, noma iziteshi zenkampani ezisemthethweni. Agcina izingxoxo eziguquguqukayo, aphendule ngokushesha, futhi aqondise isisulu aze alande ifayela elinonya noma afake iziqinisekiso kuwebhusayithi mbumbulu.
Izinga elicashile kakhulu ukwebiwa kobunikazi okusekelwe kumaphethini okuxhumana . Lapha, i-AI ifunda ukuthi umuntu wangempela ubhala kanjani kuma-imeyili akhe angaphambilini nemiyalezo: isilulumagama, isakhiwo, amagama amathegi, kanye nethoni. Umphumela uba imiyalezo ezwakala ifana ncamashi nalowo muntu, okunciphisa kakhulu ukusola komamukeli, ikakhulukazi uma engumuntu osesikhundleni segunya ngaphakathi kwenkampani.
Ezingeni elihlangene, imikhankaso yolwazi olungalungile ezenzakalelayo isebenzisa amamodeli okukhiqiza umbhalo kanye nokuhlaziya imizwa ukukhiqiza nokusabalalisa okuqukethwe okungamanga ngezinga elikhulu. Ama-bot ahlanganisiwe akhulisa imiyalezo, aphembelela ukuhlukaniswa, futhi alawule umbono womphakathi ngezindaba zezepolitiki, zempilo, noma zomnotho, ehlose ukuthonya izinqumo zomphakathi kunokumane nje antshontshe idatha eqondile.
Kube khona nezimo zokuzilungisa kwedijithali okusizwa ukuhlaziywa kwencazelo , lapho i-AI isiza khona ukuhlonza amaphrofayili asengozini (isibonelo, izingane ezincane ezinkundleni zokuxhumana) futhi ilungise inkulumo yomhlaseli ukuze yakhe ubudlelwano obuqhubekayo bokwethembana nokukhohlisa, okunengozi enkulu kakhulu ngoba kukhula kancane kancane nangokuthula.
Okokugcina, ukukhwabanisa okuhlakaniphile kusebenzisa okuqukethwe okukhiqizwe yi-AI (amadokhumenti, imibiko ecatshangelwayo, izinsiza ezikhethekile) ezenzelwe izintshisekelo zesisulu ukuze zisikhohlise ukuthi silande ifayela noma sifinyelele isayithi elinonya, sikholelwa ukuthi sisebenzisa ithuba lobuchwepheshe noma lomuntu siqu.
Ngaphandle kokuhlasela okuhlosiwe, i-AI ingadala umonakalo ngokukhetha nokungafani kwedatha yayo namamodeli. Lapho amaqembu athile emelwe kancane noma ebhalwe kabi kudatha yokuqeqesha, izinhlelo eziphumayo zivame ukukhiqiza imiphumela emibi kakhulu kulawo maqembu, zikhiqize ukucwaswa okucashile kodwa kwangempela.
Ngokwesibonelo, kwezempilo, kuye kwaphawulwa ukuthi izinhlelo zokuxilonga ezisebenzisa ikhompyutha azinembile kangako ezigulini zaseMelika Ezansundu kunasezigulini ezimhlophe ngenxa yokuntuleka kwedatha elinganiselayo. Lokhu kungaholela ekuxilongweni okubambezelekile noma okungalungile kwamaqembu athile.
Endaweni yokusebenza, amathuluzi e-AI asetshenziswa ezinqubweni zokuqasha angaqinisa imibono engajwayelekile. Into ebonakala ingenangozi njengokusebenzisa igama elithi "ninja" ekuthunyelweni komsebenzi ingadonsela iningi lamadoda, ishintshe iqembu labantu abazongenela ukhetho ngisho nangaphandle kwenhloso ecacile yokubandlulula noma ubani.
Izifundo eziphathelene namamodeli okukhiqiza izithombe zibonisa ukuthi amathuluzi amaningi akhiqiza umhlaba obandlulula ngokusobala: abaphathi cishe njalo bangamadoda amhlophe , abesifazane abavamile ukumelwa njengodokotela noma abahluleli, amadoda anesikhumba esimnyama ahlotshaniswa nobugebengu, kanye nabesifazane abanesikhumba esimnyama abanemisebenzi engaqinisekile. Ngisho noma kucelwa izithombe ezingathathi hlangothi zochwepheshe, imiphumela iveza izindima zobulili nezobuhlanga ezizuzwe njengefa kudatha.
Ohlelweni lobulungiswa bobugebengu, amathuluzi okubikezela asekelwe ku-AI avame ukuthembela kudatha yomlando yokuboshwa kanye nokuqapha. Uma leyo datha isivele iqukethe ukucwasa ngokohlanga noma ngokwendawo, amamodeli avame ukuqondisa ngokuphindaphindiwe ezindaweni kanye nemiphakathi efanayo, eqinisa amaphethini okusebenza kwamaphoyisa okungalingani futhi eqhubekisela phambili ukungalingani.
I-AI ekuphepheni kwe-inthanethi: inkemba esika nhlangothi zombili
Ngamagama okuzivikela kuphela, i-AI isibe yingxenye ebalulekile yokuphepha kwe-inthanethi kwanamuhla . Kusukela ngo-2025 nasekuqaleni kuka-2026, izimise njengethuluzi eliyisisekelo lokuhlaziya inani elikhulu lezehlakalo, ukuthola ukuziphatha okungavamile okunzima ukukubona ngeso lomuntu, ukubeka phambili izexwayiso, nokunciphisa izikhathi zokuphendula ezigamekweni.
Izinhlelo ezisekelwe ku-AI zikuvumela ukuthi uhlanganise amalogi avela emithonjeni eminingi, ubonise ukuziphatha okuvamile kwabasebenzisi namadivayisi, futhi uqalise izaziso zesikhathi sangempela lapho kutholwa ukuphambuka, okuwusizo kakhulu ezindaweni zamafu nezihlanganisiwe ezinezinkulungwane zama-endpoint.
Kodwa-ke, i-AI yokuzivikela ayithathi indawo yokwahlulela komuntu noma isidingo sezinqubo ezivuthiwe. Ngaphandle kwesu elicacile, ukuqapha okuqhubekayo, namaqembu aqeqeshwe , ukuzenzakalela kungadala umuzwa wokuphepha ongamanga noma, ngokuphambene nalokho, kuqede amandla abahlaziyi ngezexwayiso ezingabekwanga phambili.
Eminyakeni yamuva nje, abahlaseli nabo sebeqalile ukuhlanganisa i-AI ezikhalini zabo: imikhankaso yobugebengu bokweba imininingwane ebucayi ethuthukisiwe, ukuzenzisa okuthembekile, ukuqashelwa okuqondiwe okuzenzakalelayo, ukukhiqizwa kwe-malware ehlukahlukene, kanye nokulungiswa okuguquguqukayo kokuhlaselwa okusekelwe ezivikelweni abahlangabezana nazo. Umphumela uba yindawo lapho abahlaseli nabavikeli bedlala ngamathuluzi obuchwepheshe afanayo , futhi umehluko usekuqondeni nasekuzivumelaniseni nezimo.
Uma sibheka phambili, inzuzo yokuncintisana ngeke nje ibe sekusebenzisa i-AI, kodwa futhi izofakwa ngaphakathi kwesakhiwo sokuphepha esingu-360º , ngokuhlolwa okuqhubekayo, amamodeli abuyekeziwe, ukuqapha okuqhubekayo, kanye nokuphathwa okuhle kwezingozi.
Lesi sithombe sonke—kusukela ekuhlaselweni okungokoqobo njenge-GPUHammer okwenza i-GPU ingasebenzi ku-AI, kuya ekusetshenzisweni kwedatha okucashile, imikhankaso yobunjiniyela bezenhlalo eyenzelwe umuntu ngamunye, kanye nokucwasa kwesakhiwo okunomthelela emphakathini—kwenza kucace ukuthi ukuphepha kwe-AI akusadingeki; kuyisidingo esiyisisekelo sokuqhubeka nokuthembela ezinhlelweni, ngandlela thile, eziqala ukuthonya cishe zonke izinqumo ezifanele ezimpilweni zethu zedijithali futhi, ngokuya, empilweni yangempela.
