Adversarial perturbation on neural network image recognition using an evolutionary algorithm

Uspešnost prepoznavanja slik z uporabo nevronskih mrež je odvisna od parametrov in filtrov, optimiziranih skozi učni proces. Tukaj najdemo razliko v načinu prepoznavanja motivov med ljudmi in stroji. Pojavi se vrzel, ki jo napadalec s pomočjo adversarnih motenj lahko izkoristi. Slike so na videz neopazno spremenjene, ljudje razlike težko zaznajo, vendar klasifikacija nevronske mreže odpove. To delo raziskuje poustvarjanje slik z evolucijskim algoritmom. Konvolucijska nevronska mreža AlexNet po spremembi ne more prepoznati predhodno jasnih motivov. Človeku prepoznavna slika se ohrani. Pari izvirnih in poustvarjenih slik so bili primerjani z uporabo vizualne ocene in statističnih metrik.Neural networks used for image recognition heavily depend on filters and parameters optimized throughout the learning process. The difference between the way people and machines see and recognize everyday objects emerge and an attacker can use it to their advantage. The images are seemingly imperceptibly altered so that people have difficulties detecting the changes, but the classification of the neural network fails. This work explores recreating images using an evolutionary algorithm. Convolutional neural network Alexnet no longer recognizes previously clear motifs. The human recognizable image is preserved. Pairs of original and recreated images were compared using visual estimation and statistical metrics