DEEP LEARNING-BASED ABNORMAL TRAFFIC DETECTION USING BIG STEP CONVOLUTION AND ATTENTION BLOCKS

Abstract

The security and reliability of the network are affected by the discovery of unusual traffic. The primary issues that arise while attempting to detect unusual traffic are addressed by an attention-based big-step convolutional neural network traffic detection model. The similarities between features and the one-dimensional nature of the detection model are the root causes of these issues. Preprocessing the raw data into a two-dimensional grayscale picture allows us to examine the patterns of network traffic. Images with several grayscale channels are created using histogram equalization. To enhance local features, an attention approach is employed by assigning varying weights to traffic factors. When it comes to getting various traffic characteristics, pooling-free convolutional neural networks are a great way to avoid the issues that normal convolutional neural networks have, such as overfitting and missing local information. The simulation trial made use of both real-world data collecting and a diversified public dataset. We compare the proposed model against the most recent two models as well as SVM, ANN, CNN, RF, and Bayes. In a testing environment, using multiple classes resulted in an accuracy percentage of 99.5%. The proposed model is quite good at spotting issues. F1 score, accuracy, and recall are all areas where the proposed technique outperforms the status quo. There is evidence that the model is resilient and good at locating objects, even when faced with a variety of challenging scenarios.