AI-NIDS Robustness Lab
A study of how easily neural network intrusion detectors are fooled when an attacker slightly shapes its own traffic, with an interactive Streamlit dashboard.
ai-nids-robustness.streamlit.app
WHAT IT DOES
A neural network can detect network intrusions with 99.9% accuracy. But what happens when the attacker slightly reshapes its own traffic? That is what this study measures.
MLP and 1D-CNN models were trained on two datasets, the old KDD Cup 1999 and the newer UNSW-NB15, then attacked with FGSM and PGD under two threat models. Adversarial training is tested as a defence, transfer to gradient-boosted trees is measured, and decisions are explained with Integrated Gradients.
The headline: detectors that look robust on KDD break on UNSW-NB15, and the standard defence collapses there.
The Streamlit app sleeps when idle, so the first load can take a few seconds.
3.7×
UNSW-NB15 attacks sit 3.7× closer to benign traffic, so detectors break there.
HIGHLIGHTS
- Standard adversarial training drops the CNN's clean detection rate to 0.025 on UNSW-NB15.
- Mixed training avoids the collapse, but its robustness is overstated.
- Trees hold up better against transferred attacks.