Anomaly-Based Intrusion Detection System: A modern Perspective

Intrusion Detection Systems (IDS) are beneficial for providing security against malicious activities for the users. Attackers find new ways to send malicious packets without the detection of Traditional Signature Based IDS. Anomaly based intrusion detection systems using Machine Learning are hence being researched for feasibility and there are many models that have successfully classified the benchmark dataset NSL-KDD but very few research papers have used UNB-IDS 2018. Hence, the approach used is to build a model that tries to successfully classify whether it is malicious or benign with labelled packet information from both the datasets and compare them. Two datasets (A benchmark dataset and a New and a Larger Dataset) are used to train the Machine Learning Classifiers and a sufficient testing accuracy of 90 percent and above with Random Forests was achieved. The robustness of the model is tested against FGSM, JSMA and novel adversarial inputs using GANs. The model trained with NSL-KDD dataset showed significant decrease in prediction accuracy against adversarial inputs while the Model trained with UNB-IDS Dataset is robust against adversarial inputs.