Network Traffic Classification using DAE-BYOL

Full report (bit in progress)

https://docs.google.com/document/d/1UfTX9nY3h6HOpv-4OhZPO9AIa3x7MXwarfY1n4oaAeU/edit?usp=sharing

What this is

A class project testing whether combining denoising autoencoders with BYOL works for botnet detection. Got decent results on CTU-13 dataset.

The approach

1. Train a denoising autoencoder on noisy network features
2. Use the DAE encoder in a BYOL framework (in the online and target notwork)
3. Classify the learned embeddings with Random Forest

Results

• 93% accuracy on test set
• AUROC: 0.98
• Embeddings show clear separation between normal/botnet traffic