This study provides a sobering reality check for the application of Quantum Machine Learning (QML) in cybersecurity. While quantum kernel methods are often hailed as the next frontier for Network Intrusion Detection (NID), the authors demonstrate that current approaches are plagued by extreme instability and "classifier collapse"—failures that persist even when hardware noise is removed from the equation.
TL;DR
Quantum kernel methods for intrusion detection are currently outperformed by classical models like XGBoost in accuracy, speed, and stability. Crucially, the "fragility" of these quantum models isn't just a hardware noise problem; it's an intrinsic issue with how data is represented in quantum Hilbert space, leading to catastrophic failures that average accuracy metrics often hide.
What Problem Does This Paper Address?
Intrusion Detection Systems (IDS) are the "burglar alarms" of computer networks. For these systems to be useful, they must be predictable and stable; an operator needs to trust that the detection rate won't randomly plummet based on a different initialization seed.
Recent research has suggested that Quantum Kernel Methods—which map classical data into a high-dimensional quantum feature space—could improve detection. When these models underperform, researchers typically blame the "noise" (errors) inherent in current Noisy Intermediate-Scale Quantum (NISQ) devices. This paper investigates a critical question: Is hardware noise really the primary bottleneck, or is there a more fundamental problem with quantum kernels in this domain?
Methods
The authors conducted a rigorous, matched evaluation of the Pegasos-QSVC (a Quantum Support Vector Classifier) against classical heavyweights like XGBoost and RBF-SVM.
To isolate the effect of noise, they ran 56 quantum experiments across four different backends:
- Exact Noiseless Statevector: A perfect simulation with zero hardware error.
- Three Noise-Simulated Backends: (Lagos, Nairobi, and Perth) which mimic the error profiles of real IBM superconducting quantum processors.
The team used a standardized pipeline for all models, including PCA-based feature reduction to seven components and identical data splits, as shown in the experimental workflow below:

Datasets
The study utilized two industry-standard benchmarks:
- NSL-KDD: A refined version of the classic KDD Cup 99 dataset, designed to remove redundant records that bias evaluations.
- ToN-IoT: A modern dataset capturing telemetry and network traffic from a medium-scale Internet of Things (IoT) testbed.
Key Insights and Ideas
- The "Noise is the Culprit" Myth: Two of the eight observed "classifier collapses" occurred on the noiseless backend. This proves that better hardware alone won't fix the instability; the problem lies in the quantum representation itself.
- Spectral Concentration: Analysis of the Kernel Gram matrices revealed that the quantum kernels suffer from "low effective rank." Essentially, the quantum mapping squashes the data into a very small portion of the available space, making it hard for the classifier to distinguish between "normal" and "attack" traffic.
- The Danger of Averages: A model might have 90% average accuracy but fail completely on specific random seeds. In a security context, a "recall collapse" (where the model stops flagging attacks entirely) is a catastrophic failure, even if the average accuracy looks "mediocre" rather than "broken."
- Noise as a "False" Signal: Interestingly, adding noise actually increased the effective rank of the kernel matrices (see Figure 4 below). However, this wasn't "good" rank; it was just random noise flattening the spectrum without adding any useful information for classification.

Key Results
- Performance Gap: XGBoost achieved 98.94% accuracy on NSL-KDD, while the quantum model averaged only 91.80%. On the harder ToN-IoT dataset, the gap widened significantly (99.41% vs. 82.98%).
- Extreme Variance: The quantum model's performance varied by as much as 62 times more than XGBoost across different random seeds.
- Degenerate Classifiers: 14.3% of quantum runs resulted in "collapse." These fell into two camps: "Recall Collapse" (refusing to flag anything) and "Trivial-Positive" (flagging everything as an attack). These failures appear as distinct clusters away from the ideal diagonal in precision-recall space:

- Computational Cost: Training a single noisy quantum model took over 2,000 seconds, while XGBoost finished in less than half a second—a factor of nearly 4,800x speed difference.
Limitations and Future Work
The authors acknowledge that their seed counts (4 to 8 per backend) are relatively small, meaning the 14.3% failure rate is an estimate rather than a precise frequency. Additionally, they only tested one specific quantum feature map (ZFeatureMap).
Future Work: The authors call for a new reporting standard in QML research that includes worst-run performance and explicit degeneracy tests. They suggest that future research should move away from simple fidelity kernels and explore "projected" or "bandwidth-tuned" kernels that might be more resistant to the spectral concentration issues identified here.
Summary updated