# Adversarial Machine Learning in Network Intrusion Detection Systems

**Authors:** Elie Alhajjar, Paul Maxwell, Nathaniel D. Bastian  
**arXiv:** [2004.11898](https://arxiv.org/abs/2004.11898)  
**Date:** 2020-04-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.11898v1/

## Key points

- Adversarial Machine Learning in Network Intrusion Detection Systems research paper
- Explores the concept of adversarial examples designed to deceive machine learning models
- Successful use of adversarial examples in domains like image recognition, speech recognition, and spam detection
- Focus on adversarial problem in Network Intrusion Detection Systems (NIDS)
- Examination of techniques for generating adversarial examples capable of evading machine learning models
- Investigation of evolutionary computation methods and deep learning approaches for generating adversarial examples
- Evaluation using NSL-KDD and UNSW-NB15 datasets, comparing results with Monte Carlo simulation method
- Adversarial example generation techniques lead to high misclassification rates across eleven different machine learning models and a voting classifier
- Highlights vulnerability of machine learning-based NIDS to adversarial perturbations
- Importance of understanding attacks to develop robust defense mechanisms

## Summary

Adversarial Machine Learning in Network Intrusion Detection Systems is a research paper authored by Elie Alhajjar, Paul Maxwell, and Nathaniel D. Bastian. The paper explores the concept of adversarial examples, which are inputs intentionally designed by attackers to deceive machine learning models into producing incorrect outputs. Adversarial examples have been successful in various domains such as image recognition, speech recognition, and spam detection. The focus of this study is on the nature of the adversarial problem in Network Intrusion Detection Systems (NIDS). The authors specifically examine techniques used to generate adversarial examples capable of evading different machine learning models. They investigate the effectiveness of evolutionary computation methods like particle swarm optimization and genetic algorithms, as well as deep learning approaches such as generative adversarial networks for generating these adversarial examples. To evaluate the performance of these algorithms in evading NIDS, the authors apply them to two publicly available datasets: NSL-KDD and UNSW-NB15. They compare the results with a baseline perturbation method called Monte Carlo simulation. The findings reveal that their adversarial example generation techniques lead to high misclassification rates across eleven different machine learning models and a voting classifier. This research highlights the vulnerability of machine learning-based NIDS when faced with adversarial perturbations. By understanding and studying the nature of these attacks, it becomes possible to develop more robust defense mechanisms against them.