Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era

arXiv:2606.21715 · quant-ph · Submitted 2026-06-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: We're looking at the paper: "Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era".

Mira: Here's the gist of the paper. This paper investigates the effectiveness and structural limits of Variational Quantum Classifiers (VQC) for detecting network anomalies in Noisy Intermediate-Scale Quantum (NISQ) systems.

Kai: The central empirical takeaway establishes a clear directive for the field: extreme feature compression via amplitude encoding is practically inadequate for the granular detection of complex, multi-class network intrusions in the NISQ era Experimental Setup and Performance Metrics The study utilized a 4-qubit classifier featuring a 24-parameter trainable ansatz developed using amplitude encoding to embed 16 principal components On a held-out partition of the dataset, the model achieved a consistent binary-classification accuracy of 88%.

Mira: The training configuration involved using COBYLA as the primary optimizer for 500 iterations to monitor Cross-Entropy Loss.

Kai: The final evaluation on the isolated Test Set yielded an overall accuracy of 88%.

Title and authors: Kai: So, to recap this paper, it looks at Variational Quantum Classifiers, or VQCs, trying to find network anomalies in Noisy Intermediate-Scale Quantum systems <ref:2606.21715#pg1>.

Mira: Yeah, the main finding is pretty direct—they found that using amplitude encoding for feature compression isn't actually working well enough for spotting complex attacks in this NISQ era <ref:2606.21715#pg1>.

Lev: They did test it on a four qubit classifier, and while they got eighty eight percent accuracy on a smaller set of data, that performance hit a wall quickly when they tested it against the full set of attack categories <ref:2606.21715#pg1>.

Kai: The authors interpreted that plateau not as the optimizer getting stuck in a local minimum or something typical, but rather as an inherent limit to how much information you can pack into those quantum states <ref:2606.21715#pg1>.

Mira: They saw this happen when they used a class-balanced set, and instead of learning anything useful from all twenty two raw attack categories, the model just collapsed onto a small group of classes with only nine percent accuracy <ref:2606.21715#pg1>.

Lev: That collapse is serious; it means the method isn't learning to distinguish those different types of intrusions at all because the input data gets too squeezed <ref:2606.21715#pg1>.

Kai: They compared this VQC against a classical tiny multi-layer perceptron, and that classical model hit ninety seven percent accuracy under the exact same input compression, which really shows that there's a gap in what these compressed quantum states can actually capture compared to standard neural networks <ref:2606.21715#pg1>.

Mira: It means the current approach of aggressively compressing data into just sixteen principal components isn't giving us the expressiveness we need for this kind of security task <ref:2606.21715#pg1>.

Lev: So, if we stick with this amplitude encoding idea, it seems like a structural bottleneck rather than just a training problem, which is something you have to consider when planning any real hardware implementation <ref:2606.21715#pg2>.

Kai: The authors propose a major shift here—moving away from amplitude encoding and trying angle encoding instead. They suggest mapping each feature to its own dedicated qubit using rotation gates, which they say mathematically guarantees linear separability for those complex attack vectors <ref:2606.21715#pg1>.

Mira: That angle encoding idea, when combined with Cover's Theorem, is the only way they see to ensure that the quantum structure actually allows for distinguishing between those intricate attack patterns <ref:2606.21715#pg1>.

Lev: If you can guarantee that linear separability through the circuit design itself, then we might actually start seeing meaningful results instead of just hitting these structural limits <ref:2606.21715#pg2>.

Kai: They also point to a computational problem—training these complex circuits is slow because you have to run many different estimation techniques, and they suggest using dynamic low level compilation tools like JAX or PennyLane for faster gradient computations <ref:2606.21715#pg1>.

Mira: That addresses the latency issue directly; if we can make the gradients fast, we might be able to actually explore these more complex architectures on simulators sooner <ref:2606.21715#pg2>.

Lev: Faster gradient computation is crucial because even just running fifty records on this expanded sixteen qubit architecture took over ten hours in their simulation setup because it wasn't parallelizable enough yet <ref:2606.21715#pg2>.

Kai: So, what this paper really tells us is that the current way we are compressing data into quantum states isn't sufficient for complex security problems <ref:2606.21715#pg1>.

Mira: The direction they are pushing—moving to angle encoding and better compilation frameworks—is what they think is the path toward building scalable quantum machine learning architectures <ref:2606.21715#pg1>.

Lev: I think we need to see if those new architectural proposals can actually run fast enough on the hardware we have available right now, or if we still have a long way to go before that becomes practical <ref:2606.21715#pg2>.

The paper's summary: Kai: So, the authors aren't just saying this encoding strategy fails; they are proposing concrete fixes to make VQCs actually useful for things like network security <ref:2606.21715#pg1>.

Mira: Right, they’re suggesting a total change in how we map features—moving away from the amplitude encoding that caused all these issues and toward angle encoding <ref:2606.21715#pg1>.

Lev: That angle encoding idea is big because it uses rotation gates, specifically the Ry gate, to map each feature to its own independent qubit <ref:2606.21715#pg1>.

Kai: If you do that, they use Cover's Theorem on pattern separability which they claim mathematically guarantees that even complex entangled attack vectors get mapped into something linearly separable <ref:2606.21715#pg1>.

Mira: That’s a big deal because it means the structure of the circuit itself is being designed to handle those non-linear relationships better than just trying to squeeze everything into a few amplitude states <ref:2606.21715#pg1>.

Lev: They also emphasize that for training these things, we need to move toward dynamic low level compilation ecosystems like JAX or PennyLane <ref:2606.21715#pg3>.

Kai: That’s practical because it directly tackles the computational overhead of parameter shift rules, which are really what slow down the training loop when you're trying to simulate this on real hardware <ref:2606.21715#pg3>.

Mira: So, they aren't just suggesting a new way to encode data; they’re suggesting a faster way to compute the gradients for those new encodings <ref:2606.21715#pg3>.

Lev: I wonder if making the gradient computation fast enough is even feasible on current NISQ devices, or if that just pushes the problem into needing more sophisticated error correction schemes we don't have yet <ref:2606.21715#pg2>.

Kai: That’s what I’m thinking—we need to see how robust those angle encoded circuits are when you start running them on actual superconducting qubits with their inherent noise and limited coherence times <ref:2606.21715#pg2>.

The paper's improvements: Kai: So, to wrap up this paper on "Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era," it’s about showing that aggressive feature compression through amplitude encoding just doesn't cut it for spotting complex network intrusions in noisy quantum systems <ref:2606.21715#pg1>.

Mira: Exactly, and the authors clearly show that the performance ceiling they hit isn't just a simple optimization issue, but a structural limit based on how they encoded the data <ref:2606.21715#pg1>.

Lev: Yeah, so for anyone building this on real hardware right now, it means we need to be very careful about how we structure our input data before we even start worrying about the quantum error correction overhead <ref:2606.21715#pg2>.

Kai: The implication is that if you want VQCs to actually do security work in this NISQ era, you have to move toward angle encoding and better training techniques <ref:2606.21715#pg1>.

Mira: That’s the big suggestion—using rotation gates instead of amplitude encoding because it lets you map features onto independent qubits, which they say helps with linear separability <ref:2606.21715#pg1>.

Lev: From an engineering view, that shift is necessary if we want any real advantage over classical methods; it changes the architecture from something that just tries to squeeze data into a small space to something that respects the input structure more deeply <ref:2606.21715#pg2>.

Kai: They also pointed out they only tested this on a subset of the NSL-KDD dataset, which means their reported accuracy numbers are specific and not necessarily a guarantee when you try to apply this to a completely new type of attack <ref:2606.21715#pg1>.

Mira: That’s a fair caveat; you have to remember that the results are tied specifically to that twenty percent research subset they chose for testing <ref:2606.21715#pg1>.

Lev: And the computational cost is still high; they noted that processing fifty records with their expanded sixteen qubit architecture took over ten hours because it just wasn't scalable enough yet for real-world use <ref:2606.21715#pg2>.

Kai: It’s a warning, then: the theory is promising, but the hardware constraints and encoding choices are really forcing us to rethink how we build these classifiers from the ground up <ref:2606.21715#pg1>.

Mira: For anyone listening who only cares about the big picture, it shows that theoretical optimism in quantum machine learning has to be tempered by a deep understanding of how data compression affects linear separability <ref:2606.21715#pg1>.

Lev: It means our next steps aren't just tweaking optimizers; they need to focus on these fundamental changes in circuit design and how we handle those gradient computations so the work actually runs without taking forever <ref:2606.21715#pg2>.

Conclusion: Kai: So, to wrap up this paper on "Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era," we've seen that aggressive feature compression through amplitude encoding just doesn't cut it for spotting complex network intrusions in noisy quantum systems.

Mira: Exactly, and the authors clearly show that the performance ceiling they hit isn't just a simple optimization issue, but a structural limit based on how they encoded the data.

Lev: Yeah, so for anyone building this on real hardware right now, it means we need to be very careful about how we structure our input data before we even start worrying about the quantum error correction overhead.

Kai: The implication is that if you want VQCs to actually do security work in this NISQ era, you have to move toward angle encoding and better training techniques.

Mira: That’s the big suggestion—using rotation gates instead of amplitude encoding because it lets you map features onto independent qubits, which they say helps with linear separability.

Lev: From an engineering view, that shift is necessary if we want any real advantage over classical methods; it changes the architecture from something that just tries to squeeze data into a small space to something that respects the input structure more deeply.

Kai: They also pointed out they only tested this on a subset of the NSL-KDD dataset, which means their reported accuracy numbers are specific and not necessarily a guarantee when you try to apply this to a completely new type of attack.

Mira: That’s a fair caveat; you have to remember that the results are tied specifically to that twenty percent research subset they chose for testing.

Lev: And the computational cost is still high; they noted that processing fifty records with their expanded sixteen qubit architecture took over ten hours because it just wasn't scalable enough yet for real-world use.

Kai: It’s a warning, then: the theory is promising, but the hardware constraints and encoding choices are really forcing us to rethink how we build these classifiers from the ground up.

Mira: For anyone listening who only cares about the big picture, it shows that theoretical optimism in quantum machine learning has to be tempered by a deep understanding of how data compression affects linear separability.

Lev: It means our next steps aren't just tweaking optimizers; they need to focus on these fundamental changes in circuit design and how we handle those gradient computations so the work actually runs without taking forever.

Kai: So, we've seen that the current methods hit a wall due to encoding limits, but there are clear paths forward if we change the structure of how we map data onto those qubits.

Mira: It shows that theoretical promise in quantum frameworks is constrained by hardware realities, and it points toward a necessary structural change in our approach to data embedding.

Lev: We need better ways to handle the noise and the optimization hurdles before we can really use these tools for anything beyond small toy problems today.

Kai: That’s where we leave it for this paper, "Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era." It's a good reminder that in quantum machine learning, the data mapping is just as important as the training algorithm itself.

Mira: Definitely. And if we keep pushing those angle encoding ideas, we might actually start seeing some meaningful results instead of just hitting these structural limits again soon.

Lev: Next week, we’ll be looking at a paper that tries to tackle entanglement costs in non-local quantum computation; that’s a really different kind of resource management challenge.

Angelos Thomos, Theodore Andronikos

Department of Informatics, Ionian University

quant-ph

Submitted: 2026-06-19

Updated: 2026-06-19

DOI: 10.3390/jsan15040067

Code: https://github.com/Angeloth1/Quantum-IDS-NSLKDD

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 78/100

The gist: This paper investigates the effectiveness and structural limits of Variational Quantum Classifiers (VQC) for detecting network anomalies in Noisy Intermediate-Scale Quantum (NISQ) systems.

Key concepts

Amplitude Encoding
This technique maps multiple data features into the amplitudes of a quantum state. While it reduces the number of qubits needed, the paper found this method limits the model's ability to distinguish between very complex network attack patterns, causing performance to plateau.
Angle Encoding
Instead of amplitude encoding, this approach maps each feature onto an independent qubit using rotation gates. This method is proposed as a superior strategy because it leverages mathematical theorems to ensure that complex, entangled attack vectors are mapped into a state where linear separability is guaranteed.
Mode Collapse
This occurs when the optimization process causes the model to converge prematurely onto only a small subset of possible outcomes or classes. In this study, mode collapse was observed when testing finer attack categories, indicating that the excessive feature compression prevents the model from learning all necessary distinctions.
Cover's Theorem
This theorem is used to mathematically prove that if a set of vectors is sufficiently spread out in a high-dimensional space, they can be separated by a hyperplane. The authors use this principle to argue that angle encoding provides the necessary structure for complex quantum data to achieve true linear separability.

Terminology

Summary

This paper investigates the effectiveness and structural limits of Variational Quantum Classifiers (VQC) for detecting network anomalies in Noisy Intermediate-Scale Quantum (NISQ) systems. p. 0

Where the paper says it

Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era Angelos Thomos1 and Theodore Andronikos1 1 Department of Informatics, Ionian University, 7 Tsirigoti Square, 49100 Corfu, Greece; {p19thom, andronikos}@ionio. gr June 23, 2026 Abstract This paper investigates the effectiveness and structural limits of Variational Quantum Classifiers (VQC) for detecting network anomalies in the era of Noisy Intermediate-Scale Quantum (NISQ) systems.

The central empirical takeaway establishes a clear directive for the field: extreme feature compression via amplitude encoding is practically inadequate for the granular detection of complex, multi-class network intrusions in the NISQ era

Experimental Setup and Performance Metrics

The study utilized a 4-qubit classifier featuring a 24-parameter trainable ansatz developed using amplitude encoding to embed 16 principal components On a held-out partition of the dataset, the model achieved a consistent binary-classification accuracy of 88%. The training configuration involved using COBYLA as the primary optimizer for 500 iterations to monitor Cross-Entropy Loss. p. 0

Where the paper says it

, on its own training data) as a probe of representational capacity: under this configuration the VQC reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes. We interpret this behavior as consistent with the loss of linear separability induced by excessive compression in the quantum probability space, while explicitly noting that our experiments do not isolate the encoding from the ansatz depth, optimizer budget, and measurement-decoding scheme (see Limitations).

The final evaluation on the isolated Test Set yielded an overall accuracy of 88%. p. 6
Where the paper says it

The Test Set was subsequently transformed using only these locked parameters. In the final dimensionality reduction stage, Principal Component Analysis (PCA [69]) was applied to extract exactly 16 principal components.

Structural Limitations of Encoding

The authors observed that the observed performance plateau is not attributable to convergence to local minima, interpreting this as consistent with an encoding-related expressibility limit rather than an optimization artifact. The model trained on a class-balanced set spanning the 22 raw NSL-KDD attack categories and evaluated insample reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes. p. 0

Where the paper says it

As a complementary diagnostic, the model was trained on a class-balanced set spanning the 22 raw NSL-KDD attack categories and evaluated insample (i. , on its own training data) as a probe of representational capacity: under this configuration the VQC reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes.

This behavior is interpreted as consistent with the loss of linear separability induced by excessive compression in the quantum probability space. p. 0
Where the paper says it

, on its own training data) as a probe of representational capacity: under this configuration the VQC reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes. We interpret this behavior as consistent with the loss of linear separability induced by excessive compression in the quantum probability space, while explicitly noting that our experiments do not isolate the encoding from the ansatz depth, optimizer budget, and measurement-decoding scheme (see Limitations).

Comparison Against Classical Baselines

A comparative architectural parity comparison with a classical neural network, specifically a Tiny MLP with four nodes, highlighted an expressiveness gap associated with data overcompression into restricted quantum states. This classical model achieved 97% accuracy under identical input compression, significantly outperforming the VQC’s 88%. p. 2

Where the paper says it

To establish a rigorous empirical baseline, we evaluated the quantum model against a classical “Tiny MLP” with a four-node hidden layer—a strict architectural analog to the 4-qubit space. Under identical input compression, the classical model achieved 97% accuracy, exceeding the VQC’s 88%.

This divergence maps a specific methodological pathway that currently falls short, revealing that realizing the theoretical optimism of QML requires moving beyond aggressive classical dimensionality reduction and low-expressibility encodings. p. 3
Where the paper says it

maps a specific methodological pathway that currently falls short, revealing that realizing the theoretical optimism of QML requires moving beyond aggressive classical dimensionality reduction and low-expressibility encodings. • Angle Encoding as the Most Viable Path Forward.

Proposed Future Directions

The paper advocates for a paradigm shift toward scalable, high-fidelity Quantum Machine Learning architectures by proposing an alternative structure utilizing Angle Encoding. This approach maps each feature to an independent, dedicated qubit utilizing a unitary rotation gate Ry (xi). p. 8

Where the paper says it

This results in a total of 6 rotation layers (one initial, unentangled layer followed by five repetitions) applied across the 4-qubit register. Consequently, the optimization algorithm is tasked with training a vector of exactly 24 independent parameters (𝜃0 to 𝜃23).

By leveraging Cover’s Theorem on pattern separability, this method mathematically guarantees that complex, entangled attack vectors are mapped into true linear separability. p. 3
Where the paper says it

By projecting the classical feature vector into a quantum Hilbert space whose dimensionality expands exponentially with the number of qubits, we ensure that the embedded data resides in a regime where Cover’s theorem guarantees that the probability of linear separability approaches unity. In practice, this exponential lift renders even highly entangled, non-linear attack vectors overwhelmingly likely to be linearly separable.

The authors suggest future research must transition to dynamic, low-level compilation ecosystems like JAX or PennyLane to enable fast analytical gradient computations.

Dataset and Preprocessing Strategy

The NSL-KDD dataset was selected because its native 41 features strike an optimal balance between complexity and manageability for NISQ devices. Data preprocessing involved a dual-stratification strategy combining the target label with the payload size variable src bytes to mitigate distributional shift. p. 6

Where the paper says it

For data segmentation, a simple targetlabel stratification was deemed insufficient. Instead, a dual-stratification strategy was implemented, combining the target label with the payload size variable src bytes.

The final dimensionality reduction stage applied Principal Component Analysis (PCA) to extract exactly 16 principal components, which successfully preserved 86. p. 6
Where the paper says it

The Test Set was subsequently transformed using only these locked parameters. In the final dimensionality reduction stage, Principal Component Analysis (PCA [69]) was applied to extract exactly 16 principal components.

78% of the dataset’s original explained variance. p. 6
Where the paper says it

Evaluated empirically, this PCA compression successfully preserved 86. 78% of the dataset’s original explained variance.

This synergistic pipeline reduced quantum resource requirements while preserving the structural integrity of the cyberattack signatures. p. 6
Where the paper says it

78% of the dataset’s original explained variance. Ultimately, the synergistic pipeline of Target Encoding, rigorous standard scaling, and PCA-16 successfully reduced the quantum resource requirements while preserving the structural integrity of the cyberattack signatures.

Optimization Safeguards

The training process employed a dual-optimizer deployment, using COBYLA for deterministic descent and SPSA for stochastic perturbation to test the global loss landscape. The parallel evaluation conclusively demonstrated that both fundamentally distinct algorithms plateaued at exactly the same accuracy limits. p. 0

Where the paper says it

On a held-out partition of this subset, the model achieved a consistent binary-classification accuracy of 88%. A comparative evaluation with two fundamentally different optimizers, COBYLA and SPSA, found that the observed performance plateau is not attributable to convergence to local minima, a result we interpret as consistent with an encoding-related expressibility limit rather than an optimization artifact.

This behavioral symmetry provides definitive empirical proof that the 88% ceiling is not an optimization failure, but a structural consequence of Amplitude Encoding. p. 1
Where the paper says it

A negative result carries equal scientific value: by identifying precisely where and why the architecture fails, we provide the empirical constraints required to move beyond premature claims of advantage. Our experiments yield a definitive, quantitative answer to this question, mapping a specific failure mode with constructive implications.

The results also show that the model’s predictive capacity deteriorated to an overall in-sample accuracy of just 9% when confronted with the finer granularity of 22 raw NSL-KDD attack categories. p. 2
Where the paper says it

By compressing 16 features into the state space of just 4 qubits, we observe a representational bottleneck: when trained on a class-balanced set spanning the 22 raw NSL-KDD attack categories and evaluated in-sample, the model undergoes a degenerate mode collapse, with accuracy plateauing at 9%. Because the model fails to fit even the training data it was optimized on, this behaviour is suggestive of a capacity limitation rather than a pure generalization failure.

This pattern is consistent with a degenerate mode collapse where the optimizer converges to a fixed point that ignores the input entirely. p. 0
Where the paper says it

, on its own training data) as a probe of representational capacity: under this configuration the VQC reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes. We interpret this behavior as consistent with the loss of linear separability induced by excessive compression in the quantum probability space, while explicitly noting that our experiments do not isolate the encoding from the ansatz depth, optimizer budget, and measurement-decoding scheme (see Limitations).

Limitations and Validity Constraints

The authors explicitly state that the in-sample mode collapse does not by itself isolate amplitude encoding as the sole cause, as other variables like ansatz depth and optimizer budget could contribute. Furthermore, all experiments were conducted on held-out partitions of the official 20% NSL-KDD research subset, meaning reported accuracies characterize behavior within this subset rather than generalization to the full benchmark. p. 5

Where the paper says it

We emphasize that all results in this study are obtained on held-out partitions of this 20% subset; the separate official evaluation file (KDDTest+), which introduces attack types unseen during training, is deliberately outside the scope of this work due to the same simulation-cost constraints. Accordingly, our accuracy figures should be read as characterizing behaviour within the 20% subset rather than as estimates of generalization to the full NSL-KDD benchmark; evaluation on KDDTest+ is identified as a priority for future work.

The computational boundary for the expanded 16-qubit architecture exceeded 10 hours of processing for merely 50 records due to a lack of parallelizability. p. 5
Where the paper says it

Due to the extreme computational complexity of simulating Parameterized Quantum Circuits (PQCs) via classical statevector simulators, training on the entire 125,973-record dataset was deemed unfeasible. Consequently, the official 20% research subset comprising 25,192 records (KDDTrain+ 20Percent) was adopted.

The necessity of migrating to physical QPUs is highlighted as the final step to remove simulation constraints inherent to the NISQ era. p. 1
Where the paper says it

While classical machine learning has established highly effective mechanisms for handling the imbalanced distributions inherent to network traffic [3, 4], with recent deterministic studies solidifying the robustness of interpretable classical baselines like Logistic Regression on the NSL-KDD dataset [53], researchers have successfully begun translating these challenges into quantum states for diverse cybersecurity applications [6, 54, 55, 56, 57]. A major success of recent literature has been the engineering ingenuity used to deploy VQCs within strict physical NISQ limitations [58, 59].

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Replace aggressive classical dimensionality reduction with Angle Encoding to expand Hilbert space, which guarantees that complex, entangled attack vectors are mapped into true linear separability. This allows for better capture of non-linear relationships in network traffic than PCA-based compression.

  2. Implement a dynamic low-level compilation ecosystem like JAX or PennyLane to enable fast analytical gradient computations, drastically reducing the exponential overhead of parameter-shift rules. This addresses the computational bottleneck and latency issues when training complex quantum circuits.

  3. Employ a multi-seed statistical characterization for variational quantum training to account for variance, which is a natural extension given the known variance of variational quantum training. This will provide more robust empirical results than single-run simulations.

  4. Design an angle encoded VQC where each feature is mapped to an independent qubit via rotation gates, leveraging Cover’s Theorem to ensure that the probability of linear separability approaches unity. This structural change is proposed as the most viable path forward for scalable QML architectures.

Abstract

This paper investigates the effectiveness and structural limits of Variational Quantum Classifiers (VQC) for detecting network anomalies in the era of Noisy Intermediate-Scale Quantum (NISQ) systems. Using the official 20% research subset of the NSL-KDD dataset, a 4-qubit classifier featuring a 24-parameter trainable ansatz was developed, utilizing amplitude encoding to embed 16 principal components. On a held-out partition of this subset, the model achieved a consistent binary-classification accuracy of 88%. A comparative evaluation with two fundamentally different optimizers, COBYLA and SPSA, found that the observed performance plateau is not attributable to convergence to local minima, a result we interpret as consistent with an encoding-related expressibility limit rather than an optimization artifact. An architectural parity comparison with a classical neural network (a Tiny MLP with four nodes, achieving 97% accuracy) highlighted the expressiveness gap associated with data overcompression into restricted quantum states. The model was trained on a class-balanced set spanning the 22 raw NSL-KDD attack categories and evaluated in-sample as a probe of representational capacity: under this configuration the VQC reached only 9% accuracy and exhibited a degenerate mode collapse onto a small subset of classes. We interpret this behavior as consistent with the loss of linear separability induced by excessive compression in the quantum probability space, while explicitly noting that our experiments do not isolate the encoding from the ansatz depth, optimizer budget, and measurement-decoding scheme (see Limitations). Motivated by these observations and by Cover's theorem, we outline an alternative paradigm: a 16-qubit VQC with angle encoding that expands the Hilbert-space representation rather than relying on aggressive classical dimensionality reduction.

Related papers