Iqela labaphandi abavela kwiYunivesithi yasePadua (eItali) kunye neYunivesithi yaseDelft (eNetherlands) bapapashe indlela yokusebenzisa umatshini wokufunda ukuvuselela ikhowudi ye-PIN efakiweyo ukusuka kwividiyo erekhodiweyo yendawo yokufaka efakwe ngesandla ye-ATM. . Xa ufaka ikhowudi ye-PIN eneedijithi ezi-4, amathuba okuqikelela ikhowudi echanekileyo iqikelelwa kwi-41%, kuthathelwa ingqalelo ukuba kunokwenzeka ukwenza iinzame ezintathu ngaphambi kokuvimba. Kwiikhowudi ezinedijithi ezi-5 ze-PIN, uqikelelo olunokwenzeka yi-30%. Uvavanyo olwahlukileyo lwenziwa apho amavolontiya angama-78 azama ukuqikelela ikhowudi ye-PIN kwiividiyo ezirekhodiweyo ezifanayo. Kule meko, ukuba nokwenzeka kokubikezela okuyimpumelelo kwaba yi-7.92% emva kwemizamo emithathu.
Xa ugubungela iphaneli yedijithali ye-ATM ngesandla sakho, inxalenye yesandla apho igalelo lenziwe lihlala lingabonakali, elaneleyo ukuqikelela unqakrazo ngokuguqula indawo yesandla kunye nokutshintsha iminwe engagqunywanga ngokupheleleyo. Xa uhlalutya igalelo ledijithi nganye, inkqubo isusa izitshixo ezingenako ukucinezelwa kuthathelwa ingqalelo indawo yesandla sokugquma, kwaye ibala kwakhona iinketho ezinokwenzeka zokucinezela ngokusekelwe kwindawo yesandla esicinezelayo malunga nendawo yezitshixo. . Ukwandisa amathuba okufumana igalelo, isandi se-keystrokes sinokurekhodwa ukongezwa, esahluke kancinane kwisitshixo ngasinye.
Uvavanyo lusebenzise inkqubo yokufunda yomatshini esekelwe ekusetyenzisweni kwenethiwekhi ye-neural ye-convolutional (CNN) kunye nenethiwekhi ye-neural eqhubekayo esekelwe kwi-LSTM (iMemori yeXesha eliFutshane ende) ulwakhiwo. Inethiwekhi ye-CNN yayinoxanduva lokukhupha idatha yendawo kwisakhelo ngasinye, kwaye inethiwekhi ye-LSTM isebenzise le datha ukukhupha iipatheni zokwahluka kwexesha. Imodeli yaqeqeshwa kwiividiyo zabantu abahlukeneyo be-58 abafaka iikhowudi ze-PIN besebenzisa iindlela zokuvala ezikhethiweyo zabathathi-nxaxheba (umthathi-nxaxheba ngamnye wangena kwiikhowudi ezahlukeneyo ze-100, oko kukuthi, imizekelo ye-5800 yokufaka isetyenziselwa uqeqesho). Ngexesha loqeqesho, kwatyhilwa ukuba uninzi lwabasebenzisi basebenzisa enye yeendlela ezintathu eziphambili zokugubungela igalelo.
Ukuqeqesha imodeli yokufunda yomatshini, iseva esekelwe kwiprosesa ye-Xeon E5-2670 ene-128 GB ye-RAM kunye namakhadi amathathu e-Tesla K20m kunye ne-5GB yememori nganye isetyenziswe. Inxalenye yesoftware ibhalwe kwiPython isebenzisa ilayibrari yeKeras kunye neqonga leTensorflow. Ekubeni iipaneli ze-ATM zokufaka zahlukile kwaye umphumo wokuqikelela uxhomekeke kwiimpawu ezifana nobukhulu obuphambili kunye ne-topology, uqeqesho oluhlukeneyo luyafuneka kuhlobo ngalunye lwephaneli.
Njengamanyathelo okukhusela kwindlela yokuhlasela ecetywayo, kuyacetyiswa, ukuba kunokwenzeka, ukusebenzisa iikhowudi ze-PIN ezinamadijithi ama-5 endaweni ye-4, kwaye uzame ukugubungela ubuninzi besithuba sokufaka isandla ngokusemandleni akho (indlela ihlala isebenza ukuba malunga ne-75% yendawo yokufaka ihlanganiswe ngesandla sakho). Abavelisi be-ATM bacetyiswa ukuba basebenzise izikrini ezikhethekileyo ezikhuselayo ezifihla igalelo, kunye nokungabi ngoomatshini, kodwa iipaneli zokuthintela, indawo yeenombolo ezitshintsha ngokungaqhelekanga.
umthombo: opennet.ru