Annotation: Considered adversarial attacks on systems of artificial neural networks for image recognition. To increase the security of image recognition systems from adversarial attacks (avoidance attacks), the use of auto-encoders is proposed. Various attacks are considered and software prototypes of autoencoders of fully connected and convolutional architectures are developed as a means of protection against evasion attacks. The possibility of using the developed prototypes as a basis for designing autoencoders for more complex architectures is substantiated.

Keywords: image recognition system, adversarial attack, evasion attack, autoencoder.