Generative Red Teaming for Robust Network Intrusion Detection: A Constrained Adversarial Approach

The increasing reliance on digital infrastructure has made the protection of critical networks essential for ensuring operational resilience and security. Cybersecurity operations centers (CSOCs), supported by advanced technologies and skilled professionals, are central to this effort, with network intrusion detection systems (NIDS) serving as a primary defense mechanism. The adoption of machine learning (ML) and deep learning (DL) within NIDS has greatly improved their ability to detect cyber threats. However, these models are inherently vulnerable to adversarial attacks, where carefully crafted inputs manipulate model weaknesses to evade detection. The rapid evolution of generative artificial intelligence (AI) technologies has amplified this risk by enabling adversaries to automate the creation of sophisticated evasion attacks. Hence, it is critical to evaluate the robustness of ML/DL-based NIDS against evolving adversaries. Current adversarial evaluation techniques for NIDS fail to capture real-world constraints, leading to unrealistic assumptions such as unrestricted packet modifications and control over encrypted communications, resulting in a false sense of security. To bridge this gap, we introduce a generative red teaming framework designed to conduct realistic and scalable vulnerability assessment of ML/DL-based NIDS. Our adversarial framework, PerturbLite , leverages a novel autoencoder-based generative model that produces minimally perturbed, functionality-preserving adversarial network packets capable of bypassing detection. Extensive experiments on real-world benchmarking network intrusion data sets reveal that our approach achieves superior evasion performance while maintaining packet validity, often altering less than 1% of the original content. In addition to outperforming existing adversarial methods, our framework uncovers systemic vulnerabilities within NIDS, offering insights that can guide the design of more resilient detection models at a CSOC to better protect organizations. Our work advances adversarial testing methodologies in cybersecurity and aligns with best practices recommended by Cybersecurity and AI governance frameworks.

Authors

Institutions

Publication Details

Journal
ACM Transactions on Internet Technology
Published
2026-11-21
DOI
https://doi.org/10.1145/3871042
Primary Topic
Adversarial Robustness in Machine Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Generative Red Teaming for Robust Network Intrusion Detection: A Constrained Adversarial Approach

Nathaniel D. Bastian, Ankit Shah, Jalal Ghadermazi, Mohammad Noroozi et al.
ACM Transactions on Internet Technology
Adversarial Robustness in Machine Learning
article

Generative Red Teaming for Robust Network Intrusion Detection: A Constrained Adversarial Approach

Nathaniel D. Bastian, Ankit Shah, Jalal Ghadermazi, Mohammad Noroozi, Soumyadeep Hore
article en

Abstract

The increasing reliance on digital infrastructure has made the protection of critical networks essential for ensuring operational resilience and security. Cybersecurity operations centers (CSOCs), supported by advanced technologies and skilled professionals, are central to this effort, with network intrusion detection systems (NIDS) serving as a primary defense mechanism. The adoption of machine learning (ML) and deep learning (DL) within NIDS has greatly improved their ability to detect cyber threats. However, these models are inherently vulnerable to adversarial attacks, where carefully crafted inputs manipulate model weaknesses to evade detection. The rapid evolution of generative artificial intelligence (AI) technologies has amplified this risk by enabling adversaries to automate the creation of sophisticated evasion attacks. Hence, it is critical to evaluate the robustness of ML/DL-based NIDS against evolving adversaries. Current adversarial evaluation techniques for NIDS fail to capture real-world constraints, leading to unrealistic assumptions such as unrestricted packet modifications and control over encrypted communications, resulting in a false sense of security. To bridge this gap, we introduce a generative red teaming framework designed to conduct realistic and scalable vulnerability assessment of ML/DL-based NIDS. Our adversarial framework, PerturbLite , leverages a novel autoencoder-based generative model that produces minimally perturbed, functionality-preserving adversarial network packets capable of bypassing detection. Extensive experiments on real-world benchmarking network intrusion data sets reveal that our approach achieves superior evasion performance while maintaining packet validity, often altering less than 1% of the original content. In addition to outperforming existing adversarial methods, our framework uncovers systemic vulnerabilities within NIDS, offering insights that can guide the design of more resilient detection models at a CSOC to better protect organizations. Our work advances adversarial testing methodologies in cybersecurity and aligns with best practices recommended by Cybersecurity and AI governance frameworks.

ACM Transactions on Internet Technology
United States Military Academy (US), University of South Florida (US), Indiana University Bloomington (US)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon