Intelligent Anomaly Detection in Campus Networks using Lightweight Vector Quantized Conditional Weighted Wasserstein Autoencoder and Optimized Bayesian Temporal Convolutional Network

Abstract

Campus Network (CN) is a private network that provides centralized management, high-speed connectivity and secured resource sharing in larger private campuses. However, even the CN is vulnerable to security threats such as system failures, data leaks, phishing and unauthorized access due to their open and diverse environments. Recent studies have employed Machine Learning (ML) and Deep Learning (DL) algorithms for detection of complex and unidentified threats in various networks with significant success. Therefore, the DL-based methods are recommended, but the challenges of complexity, inefficient computation, high false rates and generalization to changing patterns must be handled effectively for improved anomaly detection in CN. This paper addresses these limitations by developing a hybrid model called Lightweight Vector Quantized Conditional Weighted Wasserstein Autoencoder and Optimized Bayesian Temporal Convolutional Network (LVQ-CWWAE-OBTCN).In this lightweight model, VQ-CWWAE compresses the high-dimensional traffic data into a latent representation for extracting vital features. It integrates Vector Quantization and Autoencoder with a conditional weighted loss function to extract the infrequent anomalous patterns and Wasserstein distance function for smooth clustering of the latent space representations. These latent representation features are utilized by OBTCN model consisting of Temporal Convolutional Network with the Bayesian network architecture for identifying the anomalies and alerting the CN server. Lyrebird Optimization Algorithm (LOA) is used for hyper-parameter tuning to ensure better training accuracy and computational efficiency.Evaluated on UNSW-NB15, CICIDS 2018, and SIMARGL2021 datasets, the LVQ-CWWAE-OBTCN achieved detection accuracies of 99.53%, 99.81% and 97.1%to quantify uncertainty with confidence intervals for anomaly predictions.