Ngaba ungayibuyisela umva injineli yethu yenethiwekhi ye-neural?

Usongelo olukhulayo lweNeural Network Reverse Engineering-Kwaye Ithetha ntoni kwiShishini lakho

Ngo-2024, abaphandi kwiyunivesithi enkulu babonise ukuba banokuphinda bayakhe i-architecture yangaphakathi yemodeli yolwimi olukhulu lobunikazi bengasebenzisi nto ngaphandle kweempendulo ze-API kunye ne-computing eqikelelwa kwi-2,000 yeedola. Uvavanyo luthumele ukothuka kwishishini le-AI, kodwa iziphumo zifikelela ngaphaya kweSilicon Valley. Naliphi na ishishini elihambisa iimodeli zokufunda zoomatshini - ukusuka kwiinkqubo zokubona ubuqhetseba ukuya kwiinjini zokucebisa abathengi - ngoku ujongene nombuzo ongathandekiyo: ngaba umntu angabuba ubukrelekrele ochithe iinyanga usakha? I-Neural network reverse engineering ayiseyiyo ingozi yethiyori. Yinto esebenzayo, eyanda ukufikeleleka kwisixhobo sohlaselo ekufuneka wonke umbutho oqhutywa bubuchwephesha uqondwe.

Ijongeka njani iNeural Network Reverse Engineering

Ubuyise umva ubunjineli inethiwekhi ye-neural ayifuni ufikelelo ngokwasemzimbeni kwiseva eyiqhubayo. Kwiimeko ezininzi, abahlaseli basebenzisa ubuchule obubizwa ngokuba yi-imodeli yotsalo, apho babuza ngokucwangcisiweyo i-API yomfuziselo ngegalelo elenziwe ngononophelo, emva koko basebenzise iziphumo zokuqeqesha ikopi ephantse ifane. Uphononongo luka-2023 olupapashweUSENIX Ukhuselekolubonise ukuba abahlaseli banokuphindaphinda imida yesigqibo sabahlukanisi bemifanekiso yorhwebo kunye nokuthembeka okungaphezulu kwe-95% usebenzisa imibuzo engaphantsi kwe-100,000 - inkqubo ebiza ngaphantsi kweedola ezingamakhulu ambalwa kwiintlawulo ze-API.

Ngaphandle kokutsalwa, kukhoimodeli yokuhlaselwa kwe-inversion, esebenza ngokuchaseneyo. Esikhundleni sokukopisha imodeli, abahlaseli baphinda bahlaziye idatha yoqeqesho ngokwayo. Ukuba inethiwekhi yakho ye-neural yaqeqeshwa kwiirekhodi zabathengi, izicwangciso zexabiso lobunikazi, okanye iimethrikhi zeshishini langaphakathi, uhlaselo oluyimpumelelo lwe-inversion alubi nje imodeli yakho - lubonisa idatha ebuthathaka ebhakwe kubunzima bayo. Udidi lwesithathu,ukuhlaselwa kwe-inference yobulungu, ivumela abachasi ukuba banqume ukuba indawo ethile yedatha yayiyinxalenye yesethi yoqeqesho, iphakamisa iinkxalabo ezinzulu zobumfihlo phantsi kwemimiselo efana ne-GDPR kunye ne-CCPA.

Intambo eqhelekileyo kukuba "ibhokisi elimnyama" ukucinga - ingcamango yokuba ukuhambisa imodeli emva kwe-API igcina ikhuselekile - iphukile ngokusisiseko. Lonke uqikelelo olubuyiswa yimodeli yakho yindawo yedatha umhlaseli angayisebenzisa ngokuchasene nawe.

Kutheni amaShishini kufuneka akhathalele ngaphezu kokuba awenza ngoku

Imibutho emininzi igxila kuhlahlo lwabiwo-mali lwe-cybersecurity kwi-perimeters yenethiwekhi, ukhuseleko lwe-endpoint, kunye ne-encryption yedatha. Kodwa ipropathi yengqondo efakwe kwinethiwekhi ye-neural eqeqeshiweyo inokumela iinyanga ze-R&D kunye nezigidi kwiindleko zophuhliso. Xa umntu okhuphisana naye okanye umdlali okhohlakeleyo ekhupha imodeli yakho, bafumana lonke ixabiso lophando lwakho ngaphandle kweendleko. Ngokwe-IBM's 2024 Iindleko zeNgxelo yoLwaphulo lweDatha, ukophulwa okuphakathi okubandakanya iinkqubo ze-AI kudla imibutho i-5.2 yezigidi zeedola - i-13% ephezulu kunokwaphulwa okungabandakanyi asethi ye-AI.

Umngcipheko unzima kakhulu kumashishini amancinci naphakathi. Iinkampani zamashishini zinako ukuhlawula amaqela okhuseleko azinikeleyo eML kunye neziseko zoncedo. Kodwa inani elikhulayo le-SMBs edibanisa ukufundwa koomatshini kwimisebenzi yabo - nokuba kukufumana amanqaku okukhokela, uqikelelo lwemfuno, okanye inkxaso yabathengi ezenzekelayo - ihlala ithumela iimodeli ezinobunzima obuncinci bokhuseleko. Baxhomekeke kumaqonga eqela lesithathu anokuthi aphumeze okanye angaphumezi ukhuseleko olululo.

Ingcinga eyingozi kakhulu kwi-AI yokhuseleko kukuba ubunzima bulingana nokukhuselwa. Inethiwekhi ye-neural ene-100 yezigidi zeeparameters ayikhuselekanga ngokwemvelo kunenye ene-1 yezigidi - into ebalulekileyo yindlela olawula ngayo ukufikelela kwigalelo kunye neziphumo.

Izikhuselo eziHlanu eziSebenzayo ngokuchasene noBusela obungumzekelo

Ukukhusela uthungelwano lwe-neural akufuni i-PhD kwi-adversarial machine learning, kodwa kufuna izigqibo zoyilo lwangabom. Ezi zicwangciso-qhinga zilandelayo zimele ezona ndlela zisebenzayo zincomekayo yimibutho efana ne-NIST kunye ne-OWASP ekukhuseleni iimodeli zeML ezisetyenziswayo.

• Ireyithi yokukhawulela kunye nohlahlo lwabiwo-mali lombuzo: I-Cap inani le-API libiza nawuphi na umsebenzisi omnye okanye isitshixo esinokwenza ngexesha lefestile. Uhlaselo lokutsalwa kwemodeli lufuna amashumi amawaka emibuzo — ukuncitshiswa kwesantya esinobunkunkqele kwenza ukuba utsalo olukhulu lungasebenzi ngaphandle kokuphakamisa izivusi.
• Ukuphazamiseka kwemveliso: Yongeza ingxolo elawulwayo kwimodeli yoqikelelo. Endaweni yokubuyisela amanqaku achanekileyo okuzithemba (umzekelo, 0.9237), ujikeleze ukuya kwizithuba ezinqabileyo (umzekelo, 0.92). Oku kugcina ukusetyenziswa ngelixa unyusa ngokumangalisayo inani lemibuzo umhlaseli ayifunayo ukuze akhe kwakhona imodeli yakho.
• I-Watermarking: Shicilela iisignisha ezingabonakaliyo kwindlela yokuziphatha yemodeli yakho — igalelo elithile-imveliso ngababini esebenza njengophawu lweminwe. Ukuba ikopi ebiweyo yemodeli yakho ingaphezulu, ii-watermark zibonelela ngobungqina bobusela.
• Ukwahluka kwabucala ngexesha loqeqesho: Faka ingxolo yemathematika ngexesha loqeqesho ngokwayo. Oku kuthintela ngokubonakalayo ukuba lungakanani na ulwazi malunga nawo nawuphi na umzekelo woqeqesho lomntu oluvuzayo ngoqikelelo lwemodeli, lukhusela kuzo zombini uguqulo kunye nohlaselo lwengqiqo yobulungu.
• Ukubeka iliso kunye nokubhaqwa okungaqhelekanga: Landela i-API iipateni zokusetyenziswa kweempawu zovavanyo olucwangcisiweyo. Uhlaselo lokutsalwa luvelisa usasazo lwemibuzo olwahlukileyo olujongeka kwanto njengetrafikhi esemthethweni yabasebenzisi - izilumkiso ezizenzekelayo zinokubonisa ukuziphatha okukrokrisayo ngaphambi kokuba uhlaselo luphumelele.

Ukuphumeza nokuba mibini okanye emithathu yale milinganiselo iphakamisa iindleko kunye nobunzima bokuhlaselwa ngemiyalelo yobukhulu. Injongo ayilokhuseleko olugqibeleleyo — lwenza utsalo lungabinangqondo xa kuthelekiswa nokwakha imodeli ukusuka ekuqaleni.

Indima yeZibonelelo zokuSebenza kuKhuseleko lwe-AI

Umlinganiselo omnye ongahoywanga kwiincoko malunga nokhuseleko lomfuziselo yindawo ebanzi yokusebenza. Inethiwekhi ye-neural ayikho yodwa - iqhagamshela koovimba beenkcukacha, iinkqubo zeCRM, amaqonga okuhlawula, iirekhodi zabasebenzi, kunye nezixhobo zonxibelelwano lwabathengi. Umhlaseli ongakwaziyo ukubuyisela umva injineli imodeli yakho ngokuthe ngqo angajolisa kwimibhobho yedatha eyondlayo, i-APIs etya iziphumo zayo, okanye iinkqubo zoshishino ezigcina uqikelelo lwayo.

Apha kulapho ukuba neqonga elidityanisiweyo lokusebenza kuba yinzuzo yokwenene yokhuseleko kunokuba kube lula. Xa amashishini athunga kunye uninzi lwezixhobo ze-SaaS eziqhawukileyo, indawo nganye yokudityaniswa iba yindawo yokuhlasela enokwenzeka. UMewayzijongana nale nto ngokudibanisa iimodyuli ze-207 zoshishino - ukusuka kwi-CRM kunye ne-invoyisi kwi-HR kunye nohlalutyo - kwiqonga elilodwa elinolawulo oluphakathi kunye nokungena kuphicotho. Endaweni yokukhusela izixhobo ezilishumi elinesihlanu ezahlukeneyo ezineemodeli zemvume ezilishumi elinesihlanu ezahlukeneyo, amaqela alawula yonke into esuka kwideshibhodi enye.

Kwimibutho esebenzisa amandla e-AI, oku kudityaniswa kuthetha ukunikezelwa kwedatha okumbalwa phakathi kweenkqubo, amaqhosha ambalwa e-API adadayo kwiifayile zoqwalaselo, kunye nenqaku elinye lokunyanzeliswa kwemigaqo-nkqubo yofikelelo. Xa idatha yomthengi wakho, iimethrikhi zokusebenza, kunye nengqiqo yeshishini zonke zihlala kwindawo enye elawulwayo, indawo yohlaselo yokuhluzwa kwedatha - imathiriyeli ekrwada yohlaselo lokuguqulwa kwemodeli - iyancipha kakhulu.

Iziganeko zeHlabathi zokwenyani ezitshintshe iNgxoxo

Ngo-2022, isiqalo se-fintech safumanisa ukuba umntu okhuphisana naye uphehlelele imveliso ephantse ifane yokufumana amanqaku etyala kwiinyanga nje ezisibhozo emva kokuqaliswa kokuqaliswa kwayo. Uhlalutyo lwangaphakathi lubonise ukuba umntu okhuphisana naye wayebuza ngokucwangcisiweyo i-API yokufaka amanqaku kwiinyanga, esebenzisa iimpendulo ukuqeqesha imodeli yokufanisa. Ukuqaliswa akuzange kuthintelwe izinga, kubuyise usasazo olupheleleyo olunokwenzeka, kwaye azigcinanga zilogi zemibuzo ezinokuxhasa isenzo somthetho. Umntu okhuphisana naye akazange ajongane neziphumo.

Kutshanje, ngasekupheleni kuka-2024, abaphandi bokhuseleko babonise ubuchule obubizwa ngokuba yi-"i-model ye-channel model extraction" esebenzise ukungafani kwexesha kwiimpendulo ze-API - ixesha elingakanani umncedisi uthatha ukubuyisela iziphumo zamagalelo ahlukeneyo - ukugqithisa isakhiwo sangaphakathi semodeli ngaphandle kokuhlalutya uqikelelo ngokwabo. Uhlaselo lusebenze ngokuchasene neemodeli ezibekwe kubo bobathathu ababoneleli belifu abaphambili kwaye akukho fikelelo lukhethekileyo ngaphaya kwesitshixo esiqhelekileyo se-API.

Ezi ziganeko zigxininisa inqaku elibalulekileyo:ingozi ikhula ngokukhawuleza kunokukhusela imibutho emininzi. Ubuchwephesha obabugqalwa njengophando-mda kwiminyaka emithathu eyadlulayo ngoku ziyafumaneka njengezixhobo zesixhobo esivulekileyo kwi-GitHub. Amashishini aphatha ukhuseleko lwemodeli njengenkxalabo yexesha elizayo sele esemva.

Ukwakha uKhuseleko-Okokuqala iNkcubeko ye-AI

Iteknoloji iyodwa ayizisombululi le ngxaki. Imibutho kufuneka yakhe inkcubeko apho iimpahla ze-AI ziphathwa ngokubaluleka okufanayo njengekhowudi yomthombo, iimfihlo zorhwebo, kunye nogcino-lwazi lwabathengi. Oku kuqala ngoluhlu lwempahla - iinkampani ezininzi azilugcini uluhlu olupheleleyo lokuba zeziphi iimodeli ezibekiweyo, apho zifikeleleka khona, kwaye ngubani onokufikelela kwi-API. Awunako ukukhusela into ongayaziyo ukuba ikhona.

Intsebenziswano enqamlezileyo ibalulekile. Izazinzulu zedatha kufuneka ziqonde izoyikiso ezichaseneyo. Amaqela okhuseleko kufuneka aqonde ukuba isebenza njani imibhobho yokufunda koomatshini. Abaphathi beeMveliso kufuneka benze izigqibo ezinolwazi malunga nokuba yeyiphi imodeli yolwazi evezwa yi-APIs. Ukuzilolonga rhoqo "kweqela elibomvu" - apho amaqela angaphakathi azama ukukhupha okanye ukuguqula imifuziselo yakho - abonise ubuthathaka phambi kokuba abahlaseli bangaphandle benze. Iinkampani ezifana noGoogle kunye noMicrosoft ziqhuba le mithambo qho ngekota; akukho sizathu sokuba imibutho emincinci ingakwazi ukusebenzisa iinguqulelo ezenziwe lula.

Amaqonga afana neMewayz ezisa idatha yokusebenza phantsi kophahla olunye lwenza kube lula ukunyanzelisa imigaqo-nkqubo yolawulo lwedatha echaphazela ngokuthe ngqo ukhuseleko lwe-AI. Xa ukwazi ukulandelela ukuba ngubani ofikeleleyo ukuba ngawaphi amacandelo abathengi, xa iingxelo zohlalutyo zenziwe, kunye nendlela idatha ehamba ngayo phakathi kweemodyuli, uyakha uhlobo lokuqwalaselwa okwenza zombini ukukhutshwa kwedatha okungagunyaziswanga kunye nokubiwa kwemodeli kunzima kakhulu ukuphumeza kungabonwa.

Yintoni ezayo ngokulandelayo: uLawulo, iMigangatho, kunye nokuLungisa

Ubume bolawulo buyaqhubeka. Umthetho we-EU AI Act, oye wangena kunyanzeliso ngezigaba eziqala ngo-2025, ubandakanya amalungiselelo ajikeleze imodeli elubala kunye nokhuseleko oluya kufuna ukuba imibutho ibonise ukuba ithathe amanyathelo afanelekileyo ukukhusela iinkqubo ze-AI ekuphazanyisweni nasekubiweni. E-United States, i-NIST's AI Risk Management Framework (AI RMF) ngoku ijongana ngokucacileyo nokutsalwa kwemodeli njengecandelo lezoyikiso. Amashishini azamkela ezi zicwangciso-nkqubo azakufumana kulula ukuthobela — kwaye aya kuba kwindawo engcono yokukhusela utyalo-mali lwawo lwe-AI.

Umgca osezantsi uqondile:i-neural network reverse engineering ayiyongozi eqikelelwayo egcinelwe abadlali belizwe. Bubuchwephesha obufikelelekayo, obubhalwe kakuhle ukuba nawuphi na umntu okhuphisanayo onomdla okanye umdlali okhohlakeleyo unokubusebenzisa ngokuchasene neenkqubo ezikhuselwa kakubi. Amashishini aphumeleleyo kwixesha le-AI ayizukuba ngawona akha ezona modeli zibalaseleyo - aya kuba ngawakhuselayo. Qala ngolawulo lofikelelo, ukuphazamiseka kwemveliso, kunye nokujongwa kokusetyenziswa. Yakha kwisiseko sokusebenza esimanyeneyo esinciphisa ukusasazeka kwedatha. Kwaye uphathe iimodeli zakho eziqeqeshiweyo njengempahla yexabiso eliphezulu abayiyo, kuba abo bakhuphisana nawe ngokuqinisekileyo baya kwenza njalo.

Imibuzo Ebuzwa Rhoqo

Yintoni i-neural network reverse engineering?

Neural network reverse engineering yinkqubo yokuhlalutya iziphumo zemodeli yokufunda yomatshini, iimpendulo zeAPI, okanye iipatheni zokuziphatha ukuze wakhe ngokutsha uyilo lwangaphakathi, iintsimbi, okanye idatha yoqeqesho. Abahlaseli banokusebenzisa ubuchule obufana nokutsalwa kwemodeli, ukuthelekelela ubulungu, kunye novavanyo lotshaba ukuze babe ii-algorithms zobunikazi. Kumashishini axhomekeke kwizixhobo eziqhutywa yi-AI, oku kuzisa ubunini bomgangatho ophezulu wokuqonda kunye nemingcipheko ekhuphisanayo efuna amanyathelo okhuseleko asebenzayo.

Amashishini angayikhusela njani imifuziselo ye-AI ekubeni yenziwe umva?

Ukhuseleko olungundoqo lubandakanya imibuzo ye-API, ukongeza ingxolo elawulwayo kwimodeli yeziphumo, ukujonga iipateni ezikrokrelekayo zofikelelo, kunye nokusebenzisa ubumfihlo obahlukileyo ngexesha loqeqesho. Amaqonga afana ne-Mewayz, i-OS yeshishini yeemodyuli ezingama-207, inceda iinkampani ukuba zibeke umsebenzi kwindawo enye kunye nokunciphisa ukuvezwa ngokugcina ukuhanjiswa komsebenzi we-AI obuthathaka ngaphakathi kwendawo ekhuselekileyo, ebumbeneyo kunokuba usasazeke kuyo yonke indibaniselwano yomntu wesithathu.

Ngaba amashishini amancinci asemngciphekweni wokubiwa kwemodeli ye-AI?

Ngokuqinisekileyo. Abaphandi babonise uhlaselo lokutsalwa kwemodeli oluxabisa nje i-2,000 yeedola kwikhompyutha, nto leyo ebenza ukuba bafikeleleke kuye nabani na. Amashishini amancinci asebenzisa iinjini zokucebisa ngokwesiko, i-algorithms yamaxabiso, okanye iimodeli zokubona ubuqhophololo ziithagethi ezinomtsalane ngokuchanekileyo kuba zihlala zingenalo ukhuseleko lomgangatho weshishini. Amaqonga afikelelekayo afana ne-Mewayz, aqala kwi-$19/mo ku-app.mewayz.com, nceda amaqela amancinci aphumeze ukhuseleko olusebenzayo.

Ndingenza ntoni xa ndikrokrela ukuba imodeli yam ye-AI ithotyelwe?

Qala ngokuphicotha iilogi zokufikelela kwi-API zomthamo wemibuzo engaqhelekanga okanye iipatheni zongeniso ezicwangcisiweyo ezicebisa iinzame zokutsalwa. Jikelezisa izitshixo ze-API ngoko nangoko kwaye usebenzise imida yezinga elingqongqo. Vavanya ukuba ngaba iziphumo zemodeli zivele kwiimveliso ezikhuphisanayo. Qwalasela i-watermarking kwiinguqulelo zexesha elizayo zemodeli ukulandelela ukusetyenziswa okungagunyaziswanga, kwaye udibane nengcali yokhuseleko lwe-cybersecurity ukuvavanya umda opheleleyo wolwaphulo kwaye wenze lukhuni ukhuseleko lwakho.