Adversarial Attack on ML

Overview

Machine learning firewalls and intrusion detection systems learn to block malicious network traffic from labelled examples, which makes their training data a target. This project studies data poisoning, where an attacker injects or manipulates training records to corrupt what the model learns. Using the NSL-KDD network traffic benchmark, we trained binary classifiers (normal vs. attack) on balanced data, then compared a clean model against models trained on poisoned data using three strategies: attack hiding, DoS poisoning and random label flipping.

What I did

I worked on this as part of a team. I wrote the code that prepares the data before training: poisoning it with the attack strategies, randomly flipping labels, and balancing the normal and attack classes so the models weren't biased toward the majority class. I also tested how the models performed on clean and poisoned data, analyzed the results across the different attacks, and researched how data poisoning works and how it affects detection systems.