Riaan NgoMashi 2022, kwavela ividiyo ku-inthanethi eyayibonakala ikhombisa umongameli wase-Ukraine, uVolodymyr Zelensky, ecela amasosha akhe ukuthi abeke phansi izikhali lapho ebhekene nokuhlasela kweRussia. I ividiyo— kudalwe ngosizo lwe ukuhlakanipha okungekhona okwangempela(AI)—yayimpofu ngekhwalithi futhi iqhinga lachithwa ngokushesha, kodwa njengoba okuqukethwe okokwenziwa kuba lula ukukhiqiza futhi kukholiseka ngokwengeziwe, umzamo ofanayo ngolunye usuku ungaba nemiphumela emibi kakhulu yezombangazwe.
Yingakho ngokwengxenye, njengoba ososayensi bekhompyutha bakha izindlela ezingcono zokukhiqiza ividiyo, umsindo, izithombe, nombhalo ngokwe-algorithm—imvamisa ukuze kusetshenziswe izinto ezakhayo njengokwenza amaciko akwazi ukuveza imibono yawo—bakha nama-algorithms aphikisayo ukuze bathole okuqukethwe okunjalo kokwenziwa. . Ucwaningo lwakamuva lubonisa inqubekelaphambili ekwenzeni ukutholwa kuqine kakhulu, ngezinye izikhathi ngokubheka ngale kwamasiginesha acashile amathuluzi athile esizukulwane futhi esikhundleni salokho kusetshenziswa amasignali ayisisekelo angokomzimba nawemvelo okunzima ukuthi i-AI iwalingise.
Kungenzeka futhi ukuthi okuqukethwe okukhiqizwe yi-AI nezindlela zokuhlonza zizovaleleka emuva naphambili njengoba izinhlangothi zombili ziba yinkimbinkimbi kakhulu. “Inkinga enkulu ukuthi buphathwa kanjani ubuchwepheshe obusha,” Luisa Verdolivausosayensi wamakhompyutha e- Inyuvesi yaseNaples Federico II, kusho izindlela ezintsha zokukhiqiza ezilokhu zikhula. “Kulokhu, akupheli.”
NgoNovemba, i-Intel yamemezela ukukhishwa kwayo Umtshina we-Deepfake Wesikhathi Sangempela, inkundla yokuhlaziya amavidiyo. (Igama elithi “deepfake” lisuselwa ekusetshenzisweni kokufunda okujulile—indawo ye-AI esebenzisa amanethiwekhi okwenziwa anezingqimba eziningi—ukwakha okuqukethwe okungelona iqiniso.) Cishe amakhasimende ahlanganisa izinkampani zenkundla yezokuxhumana, abasakazi, nama-NGO angasabalalisa imitshina umphakathi jikelele, uthi Ilke Demir, umcwaningi kwa-Intel. Omunye wamaphrosesa we-Intel angahlaziya ukusakazwa kwamavidiyo angama-72 ngesikhathi esisodwa. Ekugcineni le nkundla izosebenzisa amathuluzi amaningi okuhlonza, kodwa uma yethula kule ntwasahlobo izosebenzisa umtshina u-Demir awudale ngokuhlanganyela (no-Umur Çiftçi, eNyuvesi yase-Binghamton) owawubiza ngokuthi. I-FakeCatcher.
I-FakeCatcher icwaninga ngezinguquko zombala ebusweni ukuze iveze ukugeleza kwegazi, inqubo ebizwa ngokuthi i-photoplethysmography (PPG). Abacwaningi baklame isofthiwe ukuthi igxile kumaphethini athile ombala ezindaweni ezithile zobuso futhi inganaki noma yini engaphandle. Uma bebeyivumele ukuthi isebenzise lonke ulwazi olukuvidiyo, ngesikhathi sokuqeqeshwa ibingase ithembele kumasiginali amanye amajeneretha wevidiyo angawalawula kalula. “Izimpawu ze-PPG zikhethekile ngomqondo wokuthi zikuyo yonke indawo esikhunjeni sakho,” kusho uDemir. “Akuwona amehlo noma izindebe kuphela. Futhi ukushintsha ukukhanya akukuqedi, kodwa noma yikuphi ukukhiqiza kuyawaqeda, ngoba uhlobo lomsindo abawengezayo luphazamisa ukuhlobana kwendawo, i-spectral, kanye nesikhashana.” Uma ubeka enye indlela, i-FakeCatcher iqinisekisa ukuthi umbala ushintshashintsha ngokwemvelo ngokuhamba kwesikhathi njengoba inhliziyo iphampa igazi, nokuthi kunokuhlangana kuzo zonke izifunda zobuso. Ekuhlolweni okukodwa, umtshina uzuze ukunemba kwamaphesenti angama-91, cishe amaphuzu ayisishiyagalolunye angcono kunesistimu elandelayo engcono kakhulu.
Ukudalwa nokutholwa kwemidiya yokwenziwa umjaho wezikhali, lapho uhlangothi ngalunye lwakha ngakolunye. Ngokunikezwa indlela entsha yokuthola, othile ngokuvamile angaqeqesha i-algorithm yesizukulwane ukuze abe ngcono ekuyikhohliseni. Inzuzo ebalulekile ye-FakeCatcher ukuthi ayihlukaniseki, igama lezibalo elisho ukuthi ayikwazi ukuhlehliswa kalula ngenxa yokuqeqesha amajeneretha.
Ipulatifomu ye-Intel izogcina isebenzise i-a uhlelo U-Demir no-Çiftçi basanda kuthuthukiswa okuncike ekunyakazeni kobuso. Nakuba ukunyakaza kwemvelo kuthobela ukwakheka kobuso, ukunyakaza okujulile kubukeka kuhlukile. Ngakho esikhundleni sokuqeqesha inethiwekhi ye-neural kuvidiyo eluhlaza, indlela yabo kuqala isebenzisa i-algorithm yokukhulisa ukunyakaza kuvidiyo, okwenza ukunyakaza kugqame kakhulu, ngaphambi kokuyiphakela kunethiwekhi ye-neural. Ekuhlolweni okukodwa, isistimu yabo yatholwa ngokunemba okungamaphesenti angama-97 hhayi kuphela ukuthi ividiyo iyimbumbulu, kodwa iyiphi kuma-algorithms amaningana edalile, amaphuzu angaphezu kwamaphesenti amathathu angcono kunesistimu elandelayo engcono kakhulu.
Intel
Abacwaningi e- Inyuvesi yaseCalifornia eSanta Barbara yathatha indlela efanayo ephepheni lakamuva. UMichael Goebel, ongumfundi we-PhD kwezobunjiniyela bukagesi e-UCSB kanye nombhali ohlanganyela naye wephepha, uphawula ukuthi kunenhlobonhlobo yezindlela zokutholwa. “Eqinisweni, unezindlela ezingavinjelwe eziwukufunda okujulile okumsulwa,” okusho ukuthi zisebenzisa yonke idatha etholakalayo. “Ngakolunye uhlangothi, unezindlela ezenza izinto ezinjengokuhlaziya amehlo. Okwethu kuphakathi nendawo.” Uhlelo lwabo, olubizwa PhaseForensics, igxile ezindebeni futhi ikhiphe ulwazi mayelana nokunyakaza kumafrikhwensi ahlukahlukene ngaphambi kokunikeza le datha egayiwe kunethiwekhi ye-neural. “Ngokusebenzisa izici ezinyakazayo ngokwazo, senza uhlobo lwe-hardcode kokunye esifuna ukuthi inethiwekhi ye-neural ikufunde,” kusho uGoebel.
Enye inzuzo yalesi sisekelo esimaphakathi, uphawula, ukuthi i-generalization. Uma uqeqesha umtshina ongaboshiwe kumavidiyo avela kuma-algorithms esizukulwane, uzofunda ukuthola amasignesha awo kodwa hhayi lawo amanye ama-algorithms. Ithimba le-UCSB liqeqeshe i-PhaseForensics kudathasethi eyodwa, lase liyihlola kwabanye abathathu. Ukunemba kwayo kwakungamaphesenti angu-78, angu-91, namaphesenti angu-94, cishe amaphesenti amane amaphuzu angcono kunendlela yokuqhathanisa engcono kakhulu kudathasethi ngayinye.
Ama-deepfake omsindo nawo abe yinkinga. NgoJanuwari, othile walayisha umgunyathi isiqeshana yomlingisi u-Emma Watson efunda ingxenye kaHitler Mina Kampf. Nalapha futhi, abacwaningi baphezu kwecala. Ku indlela eyodwaososayensi e- Inyuvesi yaseFlorida yakha uhlelo olubonisa indlela yezwi lomuntu. Iqeqeshelwe ukurekhodwa komsindo wangempela nokungelona iqiniso, idale uhla lwamanani angokoqobo ezindaweni ezihlukene amabanga ahlukahlukene endleleni yomoya ekhiqiza umsindo. Ngokunikezwa kwesampula entsha esolisayo, inganquma ukuthi ingabe iyenzeka ngokwebhayoloji. Iphepha libika ukunemba kudathasethi eyodwa ecishe ibe ngamaphesenti angama-99.
I-algorithm yabo ayidingi ukuthi babone umsindo ojulile ovela ku-algorithm yesizukulwane esithile ukuze bavikeleke kuwo. IVerdoliva, yaseNaples, isungule enye enjalo indlela. Ngesikhathi sokuqeqeshwa, i-algorithm ifunda ukuthola amasignesha e-biometric ezikhulumi. Uma sekwenziwe, kuthatha ukurekhodwa kwangempela kwesipikha esinikeziwe, sisebenzisa amasu aso esifundiwe ukuze sithole isiginesha yebhayomethrikhi, bese sibheka leyo siginesha ekurekhodweni okungabazekayo. Kusethi yokuhlola eyodwa, ithole amaphuzu “AUC” (okucabangela amanga amanga namanegethivu angamanga) angu-0.92 kokungu-1.0. Isiqhudelana esingcono kakhulu sithole u-0.72.
Iqembu lika-Verdoliva liphinde lasebenza ekuhlonzeni izithombe ezikhiqizwe nezikhohliwe, noma zishintshwe yi-AI noma ngokusika nokunamathisela kwakudala ku-Photoshop. Baqeqeshe uhlelo olubizwa ngokuthi TruFor ezithombeni zamakhamera angu-1 475, futhi yafunda ukubona izinhlobo zokusayina ezishiywe amakhamera anjalo. Uma ubheka isithombe esisha, ingakwazi ukubona ukungafani phakathi kwamapeshi ahlukene (ngisho nakumakhamera amasha), noma isho ukuthi isithombe sonke asibukeki sengathi sivele kukhamera. Esivivinyweni esisodwa, i-TruFor ithole i-AUC engu-0.86, kuyilapho imbangi ehamba phambili yathola u-0.80. Ngaphezu kwalokho, ingagqamisa ukuthi yiziphi izingxenye zesithombe ezinomthelela omkhulu ekwahluleleni kwaso, isize abantu bahlole kabili umsebenzi waso.
Abafundi besikole samabanga aphakeme manje bavame ukuba semdlalweni wokusebenzisa i-AI ukuze benze okuqukethwe, okudala uhlelo lokukhiqiza umbhalo I-ChatGPT ukubhala izindatshana. Isixazululo esisodwa ukubuza abadali bezinhlelo ezinjalo, ezibizwa amamodeli olimi amakhulu, ukumaka umbhalo okhiqiziwe. Abacwaningi e- Inyuvesi yaseMaryland muva nje uhlongoze a indlela edala ngokungahleliwe isethi yamagama esilulumagama esohlwini oluluhlaza, bese inika ukukhetha okuncane kulawo magama lapho ibhala. Uma uwazi lolu hlu (oluyimfihlo) lwamagama asohlwini oluluhlaza, ungabheka ukuvela kwawo ocezwini lombhalo ukuze usho ukuthi mhlawumbe luvela ku-algorithm. Inkinga eyodwa ukuthi kunenombolo ekhulayo yamamodeli olimi anamandla, futhi ngeke silindele ukuthi wonke azomaka okukhiphayo.
Omunye umfundi wasePrinceton, u-Edward Tian, udale ithuluzi elibizwa ngokuthi I-GPZero ebheka izimpawu umbhalo owabhalwa ngazo I-ChatGPT ngisho ngaphandle kwe-watermarking. Abantu bathambekele ekukhetheni amagama ngendlela emangalisayo futhi baguquguquke kakhulu ngobude bomusho. Kodwa i-GPTZero ibonakala inemikhawulo. Umsebenzisi oyedwa ukubeka i-GPTZero ekuhlolweni okuncane ithole ukuthi ihlabe umkhosi ngokufanelekile imibhalo eyi-10 kweyi-10 ebhalwe nge-AI njengeyakhiwe, kodwa futhi ihlabe umkhosi ngamanga eyi-8 kweyi-10 ebhalwe ngabantu.
Ukutholwa kombhalo we-synthetic cishe kuzosalela emuva kakhulu kwezinye izindlela. Ngokuvumelana ne UTom Goldstein, uprofesa wesayensi yamakhompiyutha eNyuvesi yaseMaryland owabhala ngokuhlanganyela iphepha lokumaka, lokho kungenxa yokuthi kunokwehlukahlukana kwendlela abantu abasebenzisa ngayo ulimi, futhi ngenxa yokuthi akukho sibonakaliso esiningi. I-eseyi ingase ibe namagama angamakhulu ambalwa, uma iqhathaniswa namaphikseli ayisigidi esithombeni, futhi amagama ahlukene, ngokungafani nokuhluka okucashile kumbala wamaphikseli.
Kuningi okusengcupheni yokuthola okuqukethwe kokwenziwa. Ingasetshenziselwa ukunxenxa othisha, izinkantolo, noma abavoti. Ingakhiqiza okuqukethwe kwabantu abadala okululazayo noma okwesabisayo. Umqondo nje wama-deepfakes ungaqeda ukwethembana eqinisweni elilamule. UDemir ubiza leli kusasa ngokuthi “dystopian.” Isikhathi esifushane, uthi, sidinga ama-algorithms okubona. Isikhathi eside, sidinga izimiso eziyisisekelo ezisungula isisekelo, mhlawumbe ezibandakanya ama-watermark noma ama-blockchains.
“Abantu bangathanda ukuba nethuluzi eliwumlingo elikwazi ukwenza yonke into kahle futhi lize liyichaze,” kusho uVerdoliva mayelana nezindlela zokubona. Akukho okunjengalokho okukhona, futhi cishe kuyoke kwenzeke. “Udinga amathuluzi amaningi.” Ngisho noma umgodla wezithonjana ungehlisa ama-deepfakes, okuqukethwe kuzoba nokuphila okufushane ku-inthanethi okungenani ngaphambi kokuthi kunyamalale. Izoba nomthelela. Ngakho-ke, uVerdoliva uthi, ubuchwepheshe bubodwa abukwazi ukusisindisa. Esikhundleni salokho, abantu badinga ukufundiswa ngeqiniso elisha, elingagcwaliswanga ngokoqobo.
Kusuka Ezindaweni Zakho Zesayithi
Izihloko Ezihlobene Kuwebhu
