This article deals with the current problem of automated detection of facts of destructive influence of images on the Internet on the person. It should be noted that detecting aggressive content in images is more difficult than detecting an object and defining its category. The reason for this is that aggressive content has no specific color or object parameters, there are no common features for classification. In this case the use of convolutional neural networks with pre-learning is a successful solution. Due to increase in aggressive content on the Internet, the problem of identifying such objects in order to minimize its impact on users is acute. The paper presents the developed architecture of the neural network for solving the problem of image with aggressive content recognition. Experiments during 180 epochs showed that at the number of epochs ~ 95-100 during training and ~ during testing it is possible to achieve the classification accuracy 91.8% taking into account the top-5 results.