Autoencoder-Based Representation Learning to Predict Anomalies in Computer Networks

With the recent advances in Internet-of-thing devices (IoT), cloud-based services, and diversity in the network data, there has been a growing need for sophisticated anomaly detection algorithms within the network intrusion detection system (NIDS) that can tackle advanced network threats. Advances in Deep and Machine learning (ML) has been garnering considerable interest among researchers since it has the capacity to provide a solution to advanced threats such as the zero-day attack. An Intrusion Detection System (IDS) is the first line of defense against network-based attacks compared to other traditional technologies, such as firewall systems. This report adds to the existing approaches by proposing a novel strategy to incorporate both supervised and unsupervised learning to Intrusion Detection Systems (IDS). Specifically, the study will utilize deep Autoencoder (DAE) as a dimensionality reduction tool and Support Vector Machine (SVM) as a classifier to perform anomaly-based classification. The study diverts from other similar studies by performing a thorough analysis of using deep autoencoders as a valid non-linear dimensionality tool by comparing it against Principal Component Analysis (PCA) and tuning hyperparameters that optimizes for \u27F-1 Micro\u27 score and \u27Balanced Accuracy\u27 since we are dealing with a dataset with imbalanced classes. The study employs robust analysis tools such as Precision-Recall Curves, Average-Precision score, Train-Test Times, t-SNE, Grid Search, and L1/L2 regularization. Our model will be trained and tested on a publicly available datasets KDDTrain+ and KDDTest+