Blocking harmful images with a deep learning based next generation firewall

There are various blocking and filtering algorithms for protection against harmful contents on the Internet. However, it is impossible to classify particularly the visual contents according to their genres and block them through traditional methods. In order to block the harmful visual contents, such as various advertisements and social media posts, we need to review and classify them as per their contents. Deep learning method is today’s most efficient method to review the visual contents. In this study, only the harmful images were blocked without completely blocking the entire website. Alcoholic drinks were selected as the harmful content data set. For this purpose, a training was provided with 4.6 million images by using CNN (Convolutional Neural Networks) and GoogLeNet architecture. At the end of this training, 97.6469% of accuracy was achieved. F1 score was calculated as 87.75526188% at the end of the test conducted with 154501 images. The images were determined through the network traffic via mitmproxy and classified as harmful or harmless thanks to the trained model, and the filtering process was successfully completed.