Tawagar masu bincike daga Jami'o'in Padua (Italiya) da Delft (Netherland) sun buga hanyar yin amfani da na'ura koyo don sake ƙirƙirar PIN da aka shigar daga hoton bidiyo na wurin shigar da hannu a cikin ATM. Lokacin shigar da lambar PIN mai lamba 4, ana ƙididdige yuwuwar hasashen madaidaicin lambar a 41%, idan aka ba da yuwuwar yin ƙoƙari uku kafin toshewa. Don lambobin PIN na lambobi 5, yuwuwar hasashen shine 30%. Na dabam, an gudanar da gwaji inda masu sa kai 78 suka yi ƙoƙarin yin hasashen lambar PIN daga bidiyon da aka yi rikodi. A wannan yanayin, yuwuwar hasashe mai nasara shine 7.92% a gaban ƙoƙarin uku.
Lokacin da aka rufe na'urar dijital ta ATM tare da tafin hannu, ɓangaren hannun da ake amfani da shi don shigar da bayanai ya kasance ba a buɗe ba, wanda ya isa a iya hasashen latsawa ta hanyar canza wurin hannun da kuma canza yatsun da ba su cika ba. Lokacin nazarin shigar da kowane lambobi, tsarin ya keɓance maɓallan da ba za a iya dannawa ba, la'akari da matsayin abin rufewa, kuma yana ƙididdige mafi yawan zaɓuɓɓukan da za a iya latsawa bisa ga matsayin hannun da ake latsawa, dangane da wurin da aka latsa. makullin. Don ƙara yuwuwar tantance shigarwar, ana kuma iya rikodin sautin latsawa, wanda ya ɗan bambanta ga kowane maɓalli.
Gwajin ya yi amfani da tsarin koyo na na'ura dangane da amfani da hanyar sadarwa ta juzu'i (CNN) da kuma cibiyar sadarwa ta jijiya mai maimaitawa bisa tsarin gine-ginen LSTM (Long Short Term Memory). CNN ce ke da alhakin fitar da bayanan sararin samaniya na kowane firam, kuma cibiyar sadarwar LSTM ta yi amfani da wannan bayanan don fitar da tsarin saɓanin lokaci. An horar da samfurin akan bidiyo na mutane 58 daban-daban suna shigar da lambar PIN ta amfani da hanyoyin shigar da bayanan da mahalarta suka zaɓa (kowane ɗan takara ya shigar da lambobin 100 daban-daban, watau 5800 misalai na shigarwa an yi amfani da su don horo). A lokacin horon, an gano cewa mafi yawan masu amfani suna amfani da ɗayan manyan hanyoyin uku na shigar da bayanai.
Don horar da ƙirar koyon injin, an yi amfani da uwar garken da ke kan na’urar sarrafa Xeon E5-2670 mai 128 GB na RAM da katunan Tesla K20m uku tare da 5 GB na ƙwaƙwalwar ajiya kowanne. An rubuta ɓangaren software a cikin Python ta amfani da ɗakin karatu na Keras da dandalin Tensorflow. Tunda abubuwan shigar da ATM sun bambanta, kuma sakamakon hasashen ya dogara da halaye kamar girman da tsarin maɓalli, ana buƙatar horo daban don kowane nau'in panel.
A matsayin matakan kariya daga hanyar kai hari, ana ba da shawarar, idan zai yiwu, a yi amfani da lambobin PIN na lambobi 5 maimakon 4, sannan kuma kuyi ƙoƙarin rufe sararin shigarwa gwargwadon iko da hannunku (hanyar ta ci gaba da tasiri idan kusan 75). % na yankin shigarwa an rufe shi da hannunka). Ana ba da shawarar masana'antun ATM su yi amfani da allon kariya na musamman waɗanda ke ɓoye abubuwan shigar, da kuma ba injiniyoyi ba, amma maɓallan shigarwar taɓawa, matsayin lambobin da ke canzawa ba da gangan ba.
source: budenet.ru