msudipto/VQC-ZTI_Framework

VQC-ZTI

Variational Quantum Control for Zero Trust Protection of the Tactile Internet

Paper: VQC-ZTI: Variational Quantum Control for Zero Trust Protection of the Tactile Internet
Hugging Face Paper: huggingface.co/papers/2608.18572
Code: github.com/msudipto/VQC-ZTI_Framework
DOI: 10.48550/arXiv.2608.18572
Venue: Accepted at IEEE Global Communications Conference (GLOBECOM 2026)
Primary arXiv Category: Cryptography and Security (cs.CR)
Cross-lists: Machine Learning (cs.LG), Networking and Internet Architecture (cs.NI)

---

Model Description

VQC-ZTI is a hybrid quantum-classical anomaly-evidence model developed for a split-plane zero-trust architecture targeting Tactile Internet services.

The framework separates:

1. Off-path evidence generation, where encrypted-flow telemetry is processed by a hybrid Variational Quantum Classifier / Quantum Neural Network; and
2. On-path deterministic enforcement, where Policy Enforcement Points (PEPs) apply cached grant, restrict, step-up, or deny actions.

The design prevents probabilistic VQC inference from being placed directly in the latency-critical Tactile Internet control path.

The released implementation uses PyTorch and PennyLane and evaluates the hybrid model on CESNET-derived aggregated encrypted-flow traffic records.

Model Sources

• Paper: https://arxiv.org/abs/2608.18572
• Hugging Face Paper Page: https://huggingface.co/papers/2608.18572
• Source Code: https://github.com/msudipto/VQC-ZTI_Framework
• DOI: https://doi.org/10.48550/arXiv.2608.18572

---

Authors

Mubassir Serneabat Sudipto

Electrical and Computer Engineering
Iowa State University
Ames, Iowa, USA
Email: msudipto@iastate.edu

Shakil Ahmed

Computer Science, College of Computing
Grand Valley State University
Allendale, Michigan, USA
Email: ahmeshak@gvsu.edu

Ashfaq Khokhar

Carl R. Ice College of Engineering
Kansas State University
Manhattan, Kansas, USA
Email: akhokhar@k-state.edu

---

Model Architecture

The full-hybrid VQC-ZTI anomaly model contains the following stages:

1. Feature preprocessing

   • Payload-independent encrypted-flow statistics
   • Robust feature preparation
   • 12 input features

2. Classical embedding

   • A trainable classical embedder maps the processed feature vector into a 12-dimensional representation.

3. Quantum feature encoding

   • Number of qubits: 12
   • Initial state: all-zero state
   • Feature encoding: single-qubit Pauli-Y rotations (RY)

4. Variational Quantum Circuit

   • Number of variational layers: 2
   • Trainable single-qubit rotations
   • Nearest-neighbor CNOT entanglement

5. Quantum measurement

   • Pauli-Z expectation values are measured from two output qubits.

6. Classical classification head

   • Quantum measurements are mapped into two class logits.
   • The class-1 probability is used as the continuous anomaly score.

The complete architecture is trained end-to-end through the PyTorch-PennyLane computational graph.

---

Input

The model operates on structured, payload-independent network-traffic features derived from CESNET aggregated traffic.

The evaluated implementation uses:

• 12 numerical features
• packet and byte statistics
• destination diversity
• traffic ratios
• average flow duration
• average time-to-live
• related aggregated flow characteristics

The model is therefore a tabular binary-classification model, not a natural-language, image, or generative model.

---

Output

The model produces a continuous anomaly score:

0 <= anomaly_score <= 1

The binary experimental classes are:

0 = normative record
1 = suspicious or high-anomaly record

The score represents evidence with respect to the statistical anomaly-labeling procedure used in the study.

It must not be interpreted as a calibrated probability that a network flow is malicious.

---

Dataset and Experimental Data

The evaluation reported in the associated paper uses 4,875 CESNET-derived aggregated traffic records with 12 payload-independent features.

Experimental labels are generated through a controlled statistical-anomaly procedure based on robust-scaled feature norms and empirical quantiles.

The labeling scheme separates:

• normative records,
• suspicious records, and
• high-anomaly records.

For binary evaluation, suspicious and high-anomaly records are combined into the anomaly class.

Important Data Qualification

These labels are statistical pseudo-labels, not verified attack annotations.

Consequently:

• reported anomaly-detection performance measures agreement with the constructed statistical benchmark;
• false-positive rate is measured relative to the quantile-derived normative class; and
• the reported results must not be interpreted as verified intrusion-detection performance against confirmed cyberattacks.

The full data preparation and reproducibility workflow is documented in the GitHub repository:

https://github.com/msudipto/VQC-ZTI_Framework

---

Evaluation Protocols

Three evaluation protocols are used.

Random Stratified Holdout

• Training records: 3,900
• Evaluation records: 975
• Binary-label distribution preserved

Entity-Group Holdout

• Training records: approximately 3,751-3,971
• Evaluation records: approximately 904-1,124
• Training and evaluation entity groups are disjoint

Temporal Holdout

• Training records: 3,887 earlier records
• Evaluation records: 988 later records

These protocols test performance under conventional random splitting as well as entity and chronological distribution shifts.

---

Training Procedure

The reported full-hybrid QNN configuration uses:

Parameter	Value
Input features	12
Qubits	12
Variational layers	2
Epochs	20
Batch size	32
Learning rate	2 x 10^-3
Random seeds	42-46
Training runs	5
Quantum execution	Analytic expectation values
Quantum framework	PennyLane
ML framework	PyTorch

Training uses a class-weighted negative log-likelihood objective.

Class weights are determined from the training partition and normalized to unit mean. The classical embedder, variational quantum circuit, and classical classification head are jointly optimized.

Quantum Execution Setting

The reported experiments use noiseless analytic expectation values.

The reported results therefore do not include:

• finite-shot uncertainty,
• quantum-hardware noise,
• physical-device connectivity constraints,
• transpilation overhead,
• hardware queueing delay, or
• other physical quantum-computing effects.

---

Evaluation Results

The full-hybrid QNN is compared against:

• ExtraTrees
• Random Forest
• Logistic Regression
• hybrid-model ablations

Full-Hybrid QNN Results

Evaluation Split	ROC-AUC	Accuracy	FPR
Entity-Group	0.9974 +/- 0.0009	0.9762 +/- 0.0028	0.0241 +/- 0.0042
Random	0.9981 +/- 0.0004	0.9756 +/- 0.0042	0.0270 +/- 0.0050
Temporal	0.9941 +/- 0.0018	0.9747 +/- 0.0044	0.0248 +/- 0.0052

Across the three protocols, the full-hybrid QNN achieves the highest reported mean AUC and accuracy and the lowest mean FPR among the evaluated models.

Comparison with ExtraTrees

Relative to ExtraTrees, mean FPR is reduced by:

Evaluation Split	FPR Reduction
Random	44.6%
Entity-Group	49.6%
Temporal	67.9%

These comparisons do not demonstrate quantum computational advantage because the classical baselines are not parameter- or compute-matched to the complete hybrid neural architecture.

---

Ablation Study

The associated study evaluates five model configurations:

Variant	Description
Full Hybrid (FH)	Complete classical embedder + VQC + classification head
No Head (NH)	Classification head removed
Shallow Embedder (SE)	Reduced classical embedding architecture
No Embedder (NE)	Processed features encoded directly
PQC-only (PQC)	Restricted parameterized quantum circuit without the complete hybrid structure

The full-hybrid configuration performs best across all reported evaluation protocols.

The ablation results indicate that the reported performance depends on the complete hybrid quantum-classical pipeline, rather than on the parameterized quantum circuit alone.

---

Zero-Trust Integration

The anomaly model is designed as an evidence source, not as a direct access-control mechanism.

VQC-ZTI separates two operational planes:

Evidence Plane

The asynchronous evidence plane performs:

Encrypted-Flow Telemetry
        |
Feature Preparation
        |
Classical Embedding
        |
Quantum Encoding
        |
VQC / QNN Inference
        |
Classical Post-Processing
        |
Risk Fusion
        |
Policy Computation

Enforcement Plane

The latency-critical enforcement plane uses:

Request
   |
Policy Enforcement Point (PEP)
   |
Cached Deterministic Policy
   |
Grant / Restrict / Step-Up / Deny
   |
Protected Tactile Internet Service

VQC execution is therefore not an intermediate processing hop for the current Tactile Internet transaction.

Delayed anomaly scoring can postpone a future policy update, but does not directly add quantum-processing delay to the current enforcement decision.

---

Intended Uses

VQC-ZTI is intended for:

• research on hybrid quantum-classical machine learning;
• network anomaly-detection experimentation;
• zero-trust architecture research;
• Tactile Internet security research;
• quantum-enhanced cybersecurity experimentation;
• comparative classical/quantum model evaluation;
• reproducibility studies;
• controlled anomaly-scoring experiments;
• hybrid-model ablation studies; and
• research on off-path security-evidence generation.

---

Out-of-Scope Uses

The model should not currently be treated as:

• a production intrusion-detection system;
• a verified malware or attack detector;
• a calibrated probability-of-compromise estimator;
• an autonomous access-control authority;
• proof of quantum advantage;
• evidence of performance on physical quantum hardware;
• evidence of end-to-end Tactile Internet latency compliance; or
• a replacement for deployment-grade network-security validation.

Security-critical decisions should not be based solely on this experimental model.

---

Limitations

The evaluation has several important validity boundaries.

1. Statistical labels

   • Labels are quantile-derived statistical anomalies rather than verified attack annotations.

2. Dataset scope

   • CESNET aggregated traffic does not represent a hardware-in-the-loop Tactile Internet deployment.

3. Quantum simulation

   • Experiments use noiseless analytic quantum simulation.

4. No finite-shot evaluation

   • Shot noise and sampling uncertainty are excluded.

5. No physical quantum hardware

   • Hardware noise, topology, transpilation, execution latency, and queueing are not evaluated.

6. No demonstrated quantum advantage

   • Classical baselines are not parameter- or compute-matched to the complete hybrid architecture.

7. Zero-trust policy validation

   • Risk-fusion functions and policy thresholds have not yet been calibrated against operational access-control outcomes.

8. Latency validation

   • Prototype component measurements do not establish end-to-end Tactile Internet latency compliance.

Future work includes verified attack labels, capacity-matched neural baselines, finite-shot and noisy quantum execution, physical quantum devices, controlled attack experiments, and hardware-in-the-loop PDP/PEP evaluation.

---

Reproducibility

The complete experimental framework is available at:

GitHub:
https://github.com/msudipto/VQC-ZTI_Framework

The repository provides:

• preprocessing code;
• experiment configuration;
• PyTorch-PennyLane training workflow;
• classical baselines;
• hybrid-model ablations;
• evaluation scripts;
• result generation;
• manuscript figures;
• experiment artifacts; and
• reproducibility documentation.

Repository Installation

git clone https://github.com/msudipto/VQC-ZTI_Framework.git
cd VQC-ZTI_Framework

python -m venv .venv

On Windows:

.venv\Scripts\activate
pip install -r requirements.txt
.\run_pipeline.ps1

On compatible Unix-like environments, activate the environment with:

source .venv/bin/activate
pip install -r requirements.txt

See the GitHub repository for the current execution and data-placement instructions.

---

Downloading from Hugging Face

After replacing YOUR_HF_USERNAME with the owner of this model repository, the model repository can be cloned using:

git clone https://huggingface.co/YOUR_HF_USERNAME/VQC-ZTI

or downloaded with the Hugging Face CLI:

hf download YOUR_HF_USERNAME/VQC-ZTI

Model loading depends on the checkpoint format included with the release. Refer to the accompanying repository files and the official VQC-ZTI GitHub implementation for the exact reconstruction and inference procedure.

---

Paper

The model and experimental framework are described in:

Mubassir Serneabat Sudipto, Shakil Ahmed, and Ashfaq Khokhar.
“VQC-ZTI: Variational Quantum Control for Zero Trust Protection of the Tactile Internet.”
Accepted at IEEE Global Communications Conference (GLOBECOM 2026).
arXiv:2608.18572, 2026.

• arXiv: https://arxiv.org/abs/2608.18572
• Hugging Face Papers: https://huggingface.co/papers/2608.18572
• DOI: https://doi.org/10.48550/arXiv.2608.18572
• Code: https://github.com/msudipto/VQC-ZTI_Framework

---

Citation

If you use VQC-ZTI, its implementation, experimental methodology, model checkpoints, or reported results in academic work, please cite the associated paper.

Paper Citation

@misc{sudipto2026vqczti,
  title         = {{VQC-ZTI}: Variational Quantum Control for Zero Trust Protection of the Tactile Internet},
  author        = {Sudipto, Mubassir Serneabat and Ahmed, Shakil and Khokhar, Ashfaq},
  year          = {2026},
  eprint        = {2608.18572},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  doi           = {10.48550/arXiv.2608.18572},
  url           = {https://arxiv.org/abs/2608.18572},
  note          = {Accepted at IEEE Global Communications Conference (GLOBECOM 2026)}
}

Software Repository Citation

If you specifically use or extend the accompanying software implementation, you may additionally cite:

@misc{vqc_zti_framework_2026,
  author       = {Sudipto, Mubassir Serneabat and Ahmed, Shakil and Khokhar, Ashfaq},
  title        = {{VQC-ZTI Framework}: Variational Quantum-Classical Zero-Trust Anomaly Detection and CESNET-Based Security Evaluation},
  year         = {2026},
  howpublished = {\url{https://github.com/msudipto/VQC-ZTI_Framework}},
  note         = {Code repository},
  url          = {https://github.com/msudipto/VQC-ZTI_Framework}
}

---

License

The VQC-ZTI software implementation is released under the MIT License.

See the accompanying LICENSE file and the official GitHub repository for complete terms:

https://github.com/msudipto/VQC-ZTI_Framework

---

Acknowledgment

This research supports ongoing work in:

• quantum-enhanced cybersecurity;
• hybrid quantum-classical machine learning;
• zero-trust system design;
• network anomaly detection;
• Tactile Internet security; and
• secure next-generation network evaluation.

The project emphasizes reproducible experimental methodology and careful separation between simulated anomaly-evidence performance and claims of operational security effectiveness or quantum advantage.

---

Contact

Mubassir Serneabat Sudipto
Iowa State University
msudipto@iastate.edu

Shakil Ahmed
Grand Valley State University
ahmeshak@gvsu.edu

Ashfaq Khokhar
Kansas State University
akhokhar@k-state.edu

---

Paper: https://arxiv.org/abs/2608.18572
Hugging Face Paper: https://huggingface.co/papers/2608.18572
GitHub: https://github.com/msudipto/VQC-ZTI_Framework