Intrusion Detection via Deep Learning
Published IEEE research on network intrusion detection using deep neural networks, benchmarked against the NSL-KDD dataset.
Stack
Overview
My final-year research project at Saintgits College of Engineering, published in an IEEE conference proceeding. The goal was to see whether a relatively small neural network could outperform classical ML baselines on intrusion detection, and what trade-offs came with that.
Approach
I built and compared three models:
- A baseline random forest on engineered features
- A shallow feed-forward neural network
- A deeper architecture with dropout regularisation
Training and evaluation ran on NSL-KDD, which cleans up some of the known issues with the original KDD'99 dataset like duplicate records.
Results
The deeper network improved detection of rare attack classes (U2R and R2L) noticeably over the random forest, which is where classical models tend to struggle due to class imbalance.
The full paper covers the preprocessing pipeline, architecture, and confusion matrices across attack categories. This project is what pulled me toward working with data and, later, cloud infrastructure for model deployment.