Educational and methodological materials of the master class “Adversarial attacks on image recognition neural networks” for students and schoolchildren

The problem of neural network vulnerability has been the subject scientific research and experiments for several years. Adversarial attacks are one ways to “trick” a network, force it make incorrect classification decisions. very possibility adversarial attack lies in peculiarities machine learning networks. article shows how properties networks become source problems limitations their use. materials corresponding researches author were used as basis master class “Adversarial on image recognition networks”. presents educational class: theoretical background class, practical (in particular, single neuron is described, fast gradient sign method attacking considered), examples calculations (the uses convolutional VGG, Torch CleverHans libraries), well set typical errors students teacher’s explanations eliminate these errors. In addition, result experiment given article, its full code approbation available at above links. intended both high school university who have learned basics Python language, can also be interest computer science teachers, developers courses artificial intelligence teachers.