Intrusion Detection Systems Confront Adversarial Attacks: CNNs Lead the Charge

In the high-stakes world of network security, Intrusion Detection Systems (IDS) are the unsung heroes. These systems rely on Machine Learning (ML) to identify threats, but what happens when the models themselves become targets? Recent research has tested the resilience of popular ML architectures against adversarial attacks, turning conventional wisdom on its head.

The Experiment

Researchers evaluated three prominent architectures: a 1D Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a Random Forest (RF) ensemble. The arena? The ACI-IoT-2023 dataset, featuring over 1.2 million samples and 12 distinct attack types. The adversaries? FGSM and PGD, notorious for applying gradient-based perturbations designed to mislead ML models.

The findings were startling. Random Forest, which boasted nearly perfect baseline accuracy at 99.98%, crumbled under adversarial pressure. Even at the smallest perturbation tested, its accuracy plunged 73 percentage points. In stark contrast, CNNs maintained their composure, retaining 95.5% accuracy at a minimal perturbation level and degrading gracefully as adversarial intensity increased. LSTM's performance hovered in the middle.

What This Means for Practitioners

Why does this matter? For those deploying IDS in adversarial environments, choosing the right architecture is more than a technical choice, it's a frontline defense decision. CNNs, with their demonstrated resilience, should be the go-to architecture. This challenges the belief that high baseline accuracy is indicative of robustness. It's not just about detecting threats. it's about enduring them.

Looking Ahead

These results suggest a reevaluation of the metrics used to assess IDS effectiveness. Shouldn't we prioritize models that hold up under attack rather than those that merely perform well in benign conditions? For network security practitioners, this research provides actionable insights. CNNs aren't just preferable, they're essential for environments where adversaries are constantly refining their tactics.

The paper's key contribution: an empirical comparison of IDS architectures under adversarial conditions and a compelling case for CNNs in these roles. Code and data are available at arXiv:2606.12075v1 for those interested in further exploration.