This project aims to detect Network Intrusion of the forms Denial of Service (DoS), Probe, User to Root(U2R), and Remote to Local (R2L) using an Autoencoder + ANN Classifier model. The dataset used is NSL-KDD by University of New Brunswick.

You can run this notebook here

The done analysis done by Gerry Saporito in the article "A Deeper Dive into the NSL-KDD Data Set" , gives some insights about the structure and semantics of the dataset. The dataset has:

• 4 Categorical
• 6 Binary
• 23 Discrete
• 10 Continuous

The EDA done on this Kaggle kernel gives insights about the distribution of variables and the correlation of the features.

A custom loss function was used which is the hybrid of MSE and KL Divergence Loss, (work in progrss, yet to tune)

An exponential learning rate decay function was used:

learning_rate = initial_learning_rate * (drop ^ floor((1+epoch)/epoch_interval) )

• Modularising the code
• Combined Binary + Multiclass Classification
• Experimenting with custom layer (lambda layer for outputting absolute values)
• Experimenting with Conv1D layers for the classifier
• Multiautoencoder approach