Deep transfer learning framework for the identification of malicious activities to combat cyberattack

摘要

The people having a perpetrating mind and the facilitation in advanced technologies cause the criminogenic activities in cyberspace, thereby creating societal problems. Darknet is an internet-based technology that builds on an encrypted network. Darknet networks can be accessed using a specific software with a specific network configuration; its content does not index by any search engines. Since its beginning, Darknet has been used for criminogenic tasks and applauded primarily for cybercrime promotion, including arms and drug dealing. Few countries have control over digital media and are ruled by a suppressive government. They have formulated strict policies for freedom fighters and journalism, using the Darknet anonymously. Also, many people use it for illegal purposes. Therefore, we have both positive and negative impacts of the darknet on human society and just cannot be discarded. However, in this paper, our prime concern emanates from the darknet network detection from the network traffic data through the deep transfer learning model. To provide a more accurate result, we transform time-based features into a three-dimensional image and then feed it into a pre-trained model for the extraction of promising features. In this study, we considered the Deeplnsight method to transform the numerical features into image data. These features were then used in a proposed bi-level classification system to classify the input data into malicious activities. To identify the optimized pretrained network this paper utilized 10 pre-trained models: AlexNet, ResNetl8, ResNet50, ResNetlOl, DenseNet, GoogLeNet, VGG16, VGG19, Inceptionv3, and SqueezeNet with three different baseline classifiers, namely support vector machine, decision tree, and random forest. In addition to malicious activity prediction, the proposed model could also predict the type of traffic. The experiment results illustrate that the VGG19 based features along with random forest can classify the traffic data with 96% of accuracy.