Performance Optimization of Autoencoder Neural Network Based Model for Anomaly Detection in Network Traffic

Abstract

Network anomaly detection it is a major concern and challenging area nowadays although it provides effective and efficient mechanism from different types of attack. To enhance the security, the recent development in technological era leads to various deep learning methods for anomaly detection. Auto encoder is one of them which is better suited for network anomaly detection. In survey of existing Auto-encoder models, performance varies because there is no any specific approach mention the critical impact and importance of performance parameters and detection accuracy. To reduce model bias caused by data imbalance across different data types in the feature set, we use a new data pre-processing methodology in our proposed model that transforms and removes most affected anomaly from the sample input. Our proposed model is based on autoencoder model with five layers, employs the most effective reconstruction error function, which is critical for the model to determine whether a pattern of network traffic sample is normal or abnormal. These unique techniques, together with the ideal model architecture, enable our model to be better prepared for dimension reduction and feature learning to improve the detection accuracy and F1 score. We tested our suggested model on the CIC-IDS dataset, and it beat other similar approaches in detection, with the greatest accuracy and f1-score of 91.12 percent and 92.53 percent, respectively.