2. QShield-ZTN Architecture and Design

This section presents the QShield-ZTN framework architecture, detailing core components and design principles for quantum-resilient, autonomous 6G zero-touch networks. QShield-ZTN employs a five-layer distributed architecture: Physical Infrastructure, Quantum-Resilient Security, Zero-Trust Management, Autonomous Orchestration, and Application Service layers. The architecture integrates quantum-resistant mechanisms with AI-driven orchestration capabilities.

The multi-objective decision logic within QShield-ZTN's methodology, which jointly optimizes key network parameters subject to security and performance constraints, is presented in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Optimization and constraint processing block illustrating the multi-objective decision logic within QShield-ZTN’s methodology, where key network parameters are jointly optimized subject to security and performance constraints for adaptive orchestration in 6G environments.

https://doi.org/10.1371/journal.pone.0357590.g002

The complete five-layer distributed architecture of QShield-ZTN—comprising Physical Infrastructure, Quantum-Resilient Security, Zero-Trust Management, Autonomous Orchestration, and Application Service layers—is depicted in Fig 3.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Overall QShield-ZTN Architecture Framework showing the multi-layered approach with quantum-resilient security integration across all network domains.

https://doi.org/10.1371/journal.pone.0357590.g003

The system performance is mathematically represented as:

(1)

where represents layer performance, is network state, is security parameters, is threat vector, and are layer weights. The security framework implements post-quantum cryptographic algorithms including lattice-based cryptography, hash-based signatures, and quantum key distribution. Security strength is defined as:

(2)

where and are post-quantum and QKD security levels, is quantum entropy, and is enhancement factor.

Interpretation of Quantum Security Metric

The quantum security expression in Eq. (2) should not be interpreted as cryptographic bits-of-security in the strict cryptanalytic sense. The effective security of QShield-ZTN is governed by the minimum computational hardness among deployed primitives — the actual bound against classical and quantum adversaries. The additive entropy term does not increase cryptographic hardness beyond this minimum; rather, it serves as an orchestration-level adaptation index that weights policy decisions, trust enforcement, and resource allocation under varying threat conditions, capturing environmental uncertainty, key freshness, and channel variability. Eq. (2) is therefore a control-oriented abstraction for dynamic zero-touch orchestration, not a replacement for formal cryptographic security guarantees.

The adaptive security parameter adjustment is as follows:

(3)

where is quantum threat assessment and is adaptation rate.

Quantum key generation rate is optimized as:

(4)

where is transmission probability, is mean photon number, is binary entropy, is error rate, is error correction factor, and is block time.

Post-Quantum Cryptography Parameterization

QShield-ZTN's post-quantum layer employs a configurable LWE-based construction aligned with NIST PQC guidelines, targeting 128–256-bit security through carefully selected lattice dimensions, modulus, and Gaussian error distributions that resist both classical and quantum attacks. LWE operations are confined to session establishment and periodic key refresh phases, ensuring strong cryptographic guarantees without imposing overhead on the latency-sensitive real-time control path. The zero-trust system implements continuous verification with dynamic trust scoring:

(5)

where are trust functions, , , are behavioral, contextual, and historical features for device .

Policy enforcement decision matrix:

(6)

Micro-segmentation optimization:

(7)

where represents network segments, is communication cost, and is assignment variable. The autonomous orchestration engine follows a centralized-training, decentralized-execution (CTDE) paradigm, leveraging global state during training for stable convergence while relying on local observations at runtime for scalability and real-time responsiveness. Resource allocation is formulated as a multi-objective optimization extending classical network utility maximization to incorporate security and energy-awareness alongside performance, with adaptive priority weights enabling dynamic balancing across competing objectives under varying threat and traffic conditions. An RNN-based state predictor captures long-term temporal dependencies in traffic, threat evolution, and resource utilization, enabling proactive orchestration decisions rather than reactive mitigation — essential for zero-touch 6G environments:

(8)

where are utility functions, are priority weights, is resource allocation vector, and is total capacity.

Network state prediction using recurrent neural networks:

(9)

(10)

where is hidden state, is predicted network state, and are learnable parameters.

Autonomous decision-making reward function:

(11)

where , , are performance, security, and energy reward components.

Table 1 shows some important performance metrics at each layer of the architecture, showing the efficiency and scalability characteristics of the framework at various operational parts.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. QShield-ZTN Architecture Components and Performance Metrics.

https://doi.org/10.1371/journal.pone.0357590.t001

3.1. Simulation Setup and Experimental Configuration

To achieve reproducibility and transparency, all experiments were performed with a hybrid simulation environment combining NS-3.39 for the large-scale 6G network simulation [10], QKDNetSim for the quantum key distribution modeling [4], and TensorFlow 2.15 for the implementation of the AI-driven orchestration components [17]. The simulations were performed with a high-performance computing cluster that contains 4 Intel Xeon Platinum 8368 CPUs (3.4 GHz, 64 cores), 512 GB DDR4 memory, and 4 Nvidia A100 GPUs (80 GB HBM2). In addition, hardware-based quantum random number generator (QRNG) and Hardware Security Modules (HSMs) integration was incorporated to ensure the feasibility of real-world QKD-enabled orchestration [1].

Each experiment was performed for 300 independent runs per given scenario; for statistic consistency, synchronized random seeds were used. The simulation time was fixed at 1800 seconds for each scenario, and all the performance measures were averaged with 95% confidence intervals. To ensure fairness, we re-implemented all baseline frameworks with the reported hyperparameters and tuned them under the same conditions in order to have an unbiased comparison.

Fair Comparison Protocol

To eliminate implementation bias and ensure objective comparisons, the following fairness protocol was enforced:

Consistent Experimental Environment:

• All baselines executed in identical simulation environment (NS-3.39, QKDNetSim, TensorFlow 2.15)
• Same hardware platform (Intel Xeon Gold 6238R, NVIDIA V100, 256GB RAM)
• Identical random seed sequences for reproducibility (seeds 1–300 for 300 runs)
• Same network topology (6G testbed with 100 nodes, detailed in Section 3.1)
• Same traffic patterns and workload profiles
• Same evaluation datasets (NSL-KDD + , Edge-IIoTset, 6G-Attacks, CIC-IDS-2018)

Hyperparameter Tuning:

• All trainable baselines (AI-Enhanced, Federated Learning, Adaptive Policy, QShield-ZTN) underwent Bayesian optimization for hyperparameter selection
• Search space and budget identical across methods:
  1. Learning rate: [1e-5, 1e-2]
  2. Batch size: [64, 256, 512]
  3. Network depth: [2–5] layers
  4. Optimization budget: 100 trials per method
• Best hyperparameters selected based on validation set performance (20% of training data)
• Final evaluation on separate test set (never seen during training/tuning)

Training and Convergence:

• Maximum training time: 72 hours per method
• Convergence criterion: Rolling mean reward stable for 50 consecutive episodes
• Early stopping: Enabled with patience = 100 episodes for RL-based methods
• Checkpoint frequency: Every 50 episodes
• Best model selection: Based on validation performance

Evaluation Consistency:

• All metrics computed using identical evaluation functions (Section 3.6)
• Same measurement protocol: 300-second warm-up, 1500-second active measurement
• Statistical validation: 300 independent runs per method with different random seeds
• Confidence intervals: 95% CI reported for all stochastic metrics
• Significance testing: Paired t-tests for pairwise comparisons (α = 0.05)

Implementation Verification:

• Baseline implementations validated against published results from original papers (where available)
• Ablation studies conducted to verify component contributions
• Sanity checks performed (e.g., zero-trust baseline achieves expected behavior)

These detailed baseline implementations ensure that all reported performance comparisons are conducted under equitable conditions, with no artificial advantages provided to any method.

Table 2 presents the complete software environment and library versions used in all experiments to ensure reproducibility.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Software Environment and Library Versions.

https://doi.org/10.1371/journal.pone.0357590.t002

Latency Accounting and Cryptographic Execution Model

In QShield-ZTN, computationally intensive PQC operations (LWE-based key encapsulation, signature verification) are executed asynchronously during session initialization and key refresh phases, fully decoupled from the real-time control path. Consequently, the reported sub-millisecond latency reflects only orchestration decision-making — state collection, AI inference, policy selection, and signaling — with one-time cryptographic setup costs amortized over session lifetimes.

Hardware–Software Integration Model

NS-3 and QKDNetSim handle packet-level simulation while QRNG and HSM hardware are interfaced via API-based middleware, with cryptographic operations emulated in software and periodically validated against hardware-generated references for correctness. This intentionally decouples functional validation from cycle-accurate latency emulation, enabling scalable 6G orchestration evaluation while confirming real-world deployability.

3.2. Quantum-Resilient Cryptographic Framework

QShield-ZTN implements hybrid post-quantum cryptography combining lattice-based key exchange, hash-based signatures, and quantum key distribution. The LWE-based key generation follows:

(12)

where is random matrix, is secret key, is error vector.

Adaptive security parameter adjustment:

(13)

Quantum key rate optimization:

(14)

3.3. Zero-Trust Dynamic Policy Management

Dynamic trust scoring integrates behavioral, contextual, historical, and intelligence factors:

(15)

Policy enforcement with hysteresis prevention:

(16)

Adaptive threshold learning:

(17)

3.4. Autonomous Orchestration Algorithm

Multi-agent reinforcement learning formulation with state representation:

(18)

Multi-objective reward function:

(19)

PPO policy optimization:

(20)

Multi-agent consensus mechanism:

(21)

Algorithm 1: Integrated QShield-ZTN Orchestration and Security Algorithm

Require: Network state s₀, security constraints C, threat models MEnsure: Optimized network configuration with quantum-resilient security

1. Initialize quantum-resistant cryptographic modules

2. Initialize zero-trust policy engine with trust parameters

3. Initialize multi-agent orchestration with policy πₒ

4. Initialize threat detection ensemble

5. for episode e = 1 to E do

6. for time step t = 1 to T do

7. Collect network traffic and behavioral data

8. Update quantum threat assessment using Eq. (35)

9. Compute trust scores using Eq. (37)

10. Make policy decisions using Eq. (38)

11. Select orchestration action a_t ~ πₒ(.|s_t)

12. Execute quantum-secure communication protocols

13. Observe next state s_t+1 and compute reward using Eq. (41)

14. if threat detected OR trust score < threshold then

15. Trigger emergency response protocol

16. Update security parameters and trust thresholds

17. Implement micro-segmentation adjustments

18. end if

19. Update policy using PPO loss (Eq. 42)

20. Synchronize multi-agent states using Eq. (43)

21. Store experience for continuous learning

22. end for

23. end for

24. return Optimized quantum-resilient orchestration policy

Algorithm 1 operationalizes the proposed orchestration framework by integrating predictive state estimation, trust-aware policy enforcement, and reinforcement learning–based control within a unified decision loop. At each time step, agents observe locally available network and security signals while benefiting from globally informed policies learned during centralized training. The action selection process follows the PPO policy defined in Eq. (20), ensuring bounded policy updates and preventing abrupt control shifts that could destabilize network operation.

3.4.1 Mathematical Formulation and Implementation Mapping

The mathematical framework presented throughout this paper is directly implemented in the QShield-ZTN system architecture and experimental evaluation. This subsection provides explicit linkage between theoretical formulations and their computational realization.

Equation (1) - System Performance Function:

The multi-layer performance function defined in Equation (1):

serves as the primary evaluation metric during post-training performance assessment, not as a training objective. This function aggregates layer-wise performance contributions (physical layer, network layer, application layer, security layer, orchestration layer) to compute the overall system effectiveness score.

Implementation Details

Location in code: evaluate_system_performance() function in qshield_evaluation.py

Usage context: Called after each simulation run to compute aggregate performance

Computational procedure:

1. 1. Individual layer metrics Pᵢ are extracted from simulation logs (e.g., throughput, latency, detection accuracy)
2. 2. Metrics are normalized to [0,1] scale
3. 3. Weighted summation using layer weights w = [0.15, 0.20, 0.15, 0.30, 0.20] for physical, network, application, security, and orchestration layers respectively
4. 4. Final score P_total ∈ [0,1] reported in performance comparisons (Tables 6, 8, 10)

The weights wᵢ reflect the relative importance of each layer to overall 6G zero-touch network operation, with security and network layers assigned higher weights due to their criticality.

Equation (2) - Quantum Security Level:

The quantum security level formulation:

is used as a security strength indicator during the system initialization phase to validate cryptographic parameter selection and is continuously monitored during runtime.

Implementation Details

Location in code: quantum_security_validator() in crypto_framework.py

Usage context:

1. 1. Pre-deployment validation: Ensures selected PQC parameters (LWE dimension, modulus) and QKD configuration meet target security thresholds
2. 2. Runtime monitoring: Logs instantaneous quantum security levels based on real-time entropy measurements

Computational procedure:

1. 1. S_PQC computed from LWE hardness estimates using BKZ cost models
2. 2. S_QKD derived from current QBER and key generation rate via BB84 security analysis
3. 3. Φ_quantum measured from QRNG entropy assessments
4. 4. Enhancement factor α = 1.15 calibrated empirically
5. 5. Result validated against minimum threshold (256 bits)

This metric ensures continuous quantum resilience assurance, with alerts triggered if S_quantum drops below acceptable levels.

Equation (3) - Adaptive Security Parameter Adjustment:

The security parameter adaptation rule:

is implemented as a dynamic reconfiguration mechanism in the security layer.

Implementation Details:

• Location in code: adaptive_security_manager() in security_orchestrator.py
• Usage context: Triggered every Δt = 60 seconds or upon threat level escalation
• Computational procedure:
  1. Quantum threat assessment Q(θ) computed from threat intelligence feeds
  2. Gradient ∇Q(θ) estimated via finite differences
  3. Adaptation rate η = 0.05 (Table 5)
  4. Updated parameters θ include: encryption strength level, authentication frequency, policy enforcement strictness
  5. Changes propagated to all security enforcement points
• This equation operationalizes the zero-trust principle by continuously adjusting security posture based on threat landscape.

Equation (7) - Trust Score Calculation:

The trust scoring function:

is the core decision variable in access control policy enforcement.

Implementation Details:

• Location in code: compute_device_trust() in zero_trust_engine.py
• Usage context: Invoked for every device access request (authentication, resource allocation, communication permission)
• Computational procedure:
  1. Behavioral features B extracted from device activity logs (login patterns, data access patterns)
  2. Contextual features C derived from current environment (location, time, network state)
  3. Historical features H retrieved from device profile database
  4. Feature functions fⱼ compute normalized scores for each dimension
  5. Weighted combination using learned weights β = [0.35, 0.40, 0.25]
  6. Final trust score T ∈ [0,1] compared against dynamic threshold τ (Equation 8)
• Integration with Algorithm 1: Trust scores feed into Line 7 of Algorithm 1 for policy enforcement decisions

Equation (14) - Multi-Objective Reward Function:

The reinforcement learning reward:

is the training objective for the autonomous orchestration engine.

Implementation Details:

• Location in code: compute_reward() in rl_orchestrator.py
• Usage context: Computed at every RL timestep (Δt = 5 seconds) during training phase
• Computational procedure:
  1. Performance reward R_perf = throughput – α·latency (normalized)
  2. Security reward R_sec = detection_rate – β·false_positive_rate (normalized)
  3. Energy reward R_energy = -total_power_consumption (normalized)
  4. Weights: w_p = 0.4, w_s = 0.5, w_e = 0.1 (Table 5)
  5. Combined reward R used in PPO policy gradient update (Equation 19)
• Training convergence: This reward maximization drives the learning process shown in Fig 7(a)

Equation (19) - PPO Policy Optimization:

The Proximal Policy Optimization update:

is the optimization algorithm for training the orchestration policy network.

Implementation Details:

• Location in code: ppo_update() in rl_training.py
• Usage context: Applied every K = 2048 timesteps during training (Table 5)
• Computational procedure:
  1. Collect trajectory data (states, actions, rewards, advantages) over K timesteps
  2. Compute probability ratios r(θ) = π_new(a|s) / π_old(a|s)
  3. Clip ratios to [1-ε, 1 + ε] with ε = 0.2
  4. Calculate clipped objective L^CLIP
  5. Perform M = 10 gradient ascent mini-batch updates
  6. Update policy network parameters θ
• Experimental validation: Convergence behavior shown in Table 12 and Fig 7

Equation (22) - Resource Allocation Optimization:

The resource allocation problem:

is solved as a real-time decision problem during network operation.

Implementation Details:

• Location in code: optimize_resource_allocation() in resource_manager.py
• Usage context: Invoked when new resource requests arrive or network state changes
• Computational procedure:
  1. Formulate constrained optimization with current demands and available capacity C
  2. Security constraints S_constraints enforce minimum security levels per device
  3. Solved using projected gradient descent with 100 iterations
  4. Allocation vector r applied to network resource provisioning
  5. Performance monitored via resource utilization efficiency (RUE metric)

The consensus update in Eq. (21) enables coordination among distributed agents without requiring centralized synchronization during execution. This mechanism ensures that local decisions converge toward globally consistent orchestration behavior, which is particularly important in large-scale 6G deployments spanning heterogeneous domains such as terrestrial, aerial, and satellite networks.

Table 3 details the neural network architecture specifications for all learnable components within QShield-ZTN. The policy and value networks follow a standard actor-critic architecture with three fully-connected layers. The RNN state predictor uses a two-layer LSTM architecture with 256 hidden units per layer to capture temporal dependencies in network state evolution. The anomaly detector employs a deep neural network classifier for real-time threat discrimination. The total trainable parameter count of approximately 1.7M is deliberately kept moderate to ensure feasibility on edge computing devices and to minimize inference latency during real-time orchestration.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Neural network architecture specifications.

https://doi.org/10.1371/journal.pone.0357590.t003

Overall, the proposed learning and orchestration strategy extends existing reinforcement learning–based network control approaches by embedding quantum-resilient security awareness directly into the state, reward, and policy update mechanisms, rather than treating security as an external constraint.

3.5. Performance Optimization and Threat Detection

Resource allocation optimization with security constraints:

(22)

where utility function incorporates security costs:

Ensemble threat detection:

(23)

Multi-modal anomaly detection:

(24)

3.6. Experimental Setup and Datasets

The evaluation utilizes hybrid simulation environment with NS-3 extended quantum modules, Intel Xeon processors with HSMs, NVIDIA GPUs, and quantum random number generators. Multiple datasets ensure comprehensive assessment:

To ensure a comprehensive evaluation of QShield-ZTN, we used four heterogeneous datasets covering diverse 6G security and orchestration scenarios. The NSL-KDD+ dataset(https://www.kaggle.com/datasets/hassan06/nslkdd) (148,517 samples, 43 features) was used to benchmark classical intrusion detection performance, while the CICIDS2017 + dataset (https://www.kaggle.com/datasets/chethuhn/network-intrusion-dataset) (2.83M samples, 78 features) provided a high-dimensional testing environment covering 14 attack categories. To evaluate zero-touch orchestration resilience and multi-domain attack mitigation, we developed the QShield-Synthetic dataset, comprising 5.26M samples and 156 features, modeling dynamic slice allocation, orchestration delays, cross-domain failures, and multi-vector cyberattacks. Finally, the Quantum-Dataset (https://www.kaggle.com/datasets/bhagvendersingh/quantum-dataset) (850K samples, 67 features) simulated QKD channel impairments, entanglement losses, and quantum eavesdropping attacks.

All datasets were preprocessed using a unified pipeline to ensure fairness. Continuous attributes were normalized using min-max scaling, categorical features were transformed via one-hot encoding, and missing values were imputed using an iterative KNN-based approach. Hyperparameters of QShield-ZTN were optimized using a Bayesian Optimization strategy combined with 5-fold cross-validation, exploring learning rates , policy update intervals , and reward-balancing coefficients . Each baseline model was retrained on the same datasets using its reported optimal configurations to ensure unbiased performance comparisons.

Table 4 presents experimental datasets characteristics and key performance metrics comparison, demonstrating QShield-ZTN’s superior performance across all evaluation criteria.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Experimental datasets and performance metrics.

https://doi.org/10.1371/journal.pone.0357590.t004

Hardware-in-the-loop integration with QRNG and HSM devices is employed for functional validation of cryptographic primitives and entropy sources, while timing-critical performance metrics are evaluated using software-level abstractions to ensure scalability and repeatability.

3.6.1. Performance Metrics Definition and Calculation.

The performance evaluation of QShield-ZTN is based on rigorously defined metrics across security, network performance, orchestration efficiency, and quantum resilience dimensions. Each metric is formally defined below with its calculation methodology and measurement protocol.

Security performance metrics:

Detection Accuracy (DA): Detection accuracy quantifies the ratio of correctly identified instances (both threats and benign traffic) to the total number of instances evaluated. Formally:

where TP denotes true positives (correctly detected threats), TN denotes true negatives (correctly classified benign traffic), FP denotes false positives (benign traffic misclassified as threats), and FN denotes false negatives (threats misclassified as benign). This metric is computed over the entire test dataset comprising 29,703 samples as detailed in Table 4.

False Positive Rate (FPR) FPR measures the proportion of benign instances incorrectly flagged as threats:

Lower FPR values indicate reduced operational burden from false alarms. The target threshold is maintained below 5% to ensure practical deployment feasibility.

Precision and Recall Precision quantifies the proportion of detected threats that are genuine:

Recall (also termed sensitivity or true positive rate) measures the proportion of actual threats successfully detected:

The F1-score provides the harmonic mean of precision and recall:

Quantum Resistance Metrics

Quantum Security Strength (QSS): The quantum security strength is defined in terms of computational hardness against quantum adversaries, measured in equivalent classical security bits. For lattice-based constructions, this is determined by:

where Cost_classical represents the classical attack complexity (e.g., BKZ lattice reduction) and Cost_quantum represents quantum attack complexity (e.g., Grover's algorithm applied to lattice problems). QShield-ZTN achieves QSS ≥ 256 bits through LWE parameters: dimension n = 1024, modulus q = 2²⁷, and error distribution χ with standard deviation σ = 8.

The reported “quantum resistance improvement” refers to the relative increase in QSS compared to classical cryptographic schemes vulnerable to quantum attacks (e.g., RSA-2048 or ECC-256, which offer 0 bits of quantum security):

For non-quantum-resistant baselines, this is reported as achieving >256-bit quantum security versus 0-bit quantum security.

Quantum Key Distribution Rate (QKDR): QKDR measures the rate of secure key generation through QKD channels, expressed in bits per second:

where Y₀ is the single-photon detection probability, μ is the mean photon number per pulse, f(e) is the efficiency factor dependent on quantum bit error rate (QBER) e, and T_block is the temporal block duration. Measured QKDR values for QShield-ZTN range from 0.8 to 1.5 MHz under realistic channel conditions with QBER ≤ 3.5%.

Network Performance Metrics

Throughput: Network throughput is measured as the aggregate data successfully transmitted per unit time:

Measurements are conducted under steady-state conditions with 1000 concurrent network flows, reported in Gbps. The cryptographic overhead impact is isolated by comparing throughput with and without security enforcement active.

End-to-End Latency: Latency quantifies the time interval from packet generation at the source to successful reception at the destination:

This includes propagation delay, transmission delay, queueing delay, and processing delay (including security operations during steady-state operation, excluding initial handshake). Reported values represent the mean across 10,000 packet transmissions with 95% confidence intervals.

Packet Loss Rate (PLR): PLR is the fraction of packets lost during transmission:

Lower PLR indicates improved network reliability. Target PLR < 0.1% for mission-critical 6G applications.

Energy Efficiency (EE): Energy efficiency quantifies the data transmitted per unit energy consumed:

Expressed in bits/Joule, this metric accounts for computational energy (cryptographic operations, ML inference) and transmission energy. Measurements utilize hardware power monitoring over 1800-second simulation intervals.

Orchestration Efficiency Metrics

Orchestration Decision Latency (ODL) ODL measures the time required for the autonomous orchestration engine to process network state and generate control decisions:

This excludes cryptographic session setup time and reflects only the RL policy inference latency. Reported as mean ± standard deviation over 50,000 orchestration decisions.

Convergence Episodes: The number of training episodes required for the reinforcement learning policy to achieve stable performance, defined as the episode count when the rolling mean reward (window size = 50) first exceeds 95% of the asymptotic maximum and remains stable thereafter.

Resource Utilization Efficiency (RUE): RUE quantifies the ratio of successfully allocated resources to total available resources:

Higher values indicate efficient resource usage with minimal waste.

Statistical Validation Methodology

All performance metrics are evaluated across 300 independent simulation runs with different random seeds to ensure statistical robustness. Results are reported as mean ± standard deviation. Statistical significance is assessed using paired t-tests (α = 0.05) for pairwise comparisons and one-way ANOVA (α = 0.05) for multi-group comparisons, as detailed in Section 5.3.

Measurement Protocol

Performance data collection follows a standardized protocol:

• System initialization with specified parameters (Table 5)
• Warm-up period of 300 seconds (excluded from measurements)
• Active measurement period of 1500 seconds
• Data logging at 1-second intervals
• Post-processing to compute aggregate metrics

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Complete Hyperparameter Configuration of QShield-ZTN.

https://doi.org/10.1371/journal.pone.0357590.t005

This protocol ensures reproducibility and minimizes transient effects on reported performance.

Model training was conducted using five-fold cross-validation on the training set, while the validation set was reserved exclusively for hyperparameter selection.

Table 5 presents the complete hyperparameter configuration used in all QShield-ZTN experiments. Parameters marked ‘Fixed’ were set based on established best practices in reinforcement learning and network security literature, while those marked ‘Bayesian Optimization’ were selected through the automated hyperparameter tuning procedure described in Section IV-C. The LWE cryptographic parameters follow NIST post-quantum cryptography guidelines for 128–256-bit security levels. All configurations were applied consistently across the 300 independent simulation runs to ensure reproducibility.

A full reproducibility package for the QShield-ZTN framework is publicly available at https://github.com/Sulaimanamro/QZT under the MIT License. The repository contains the complete source code for all framework components (PPO orchestration engine, RNN state predictor, zero-trust policy engine, threat detection module, and quantum-resilient cryptographic wrappers), along with all hyperparameter configurations, random seeds, and Bayesian optimization. NS-3.39 simulation scripts and automated orchestration tools reproduce all 300 independent runs per scenario, while re-implemented baselines ensure fair comparison reproducibility. Public datasets (NSL-KDD + , CICIDS2017 + , Quantum-Dataset) are linked via Kaggle, and the QShield-Synthetic generation script is included with configurable parameters. Evaluation scripts cover all reported metrics and statistical tests (ANOVA, t-tests, Wilcoxon), and a hardware abstraction layer supports both simulated and hardware-in-the-loop execution. Full setup and reproduction instructions are provided in the README with pinned Python dependencies.

4.1. Security Performance Results

The security performance evaluation demonstrates the superior effectiveness of QShield-ZTN in protecting against both classical and quantum computing threats. The quantum resistance strength achieved by QShield-ZTN significantly exceeds that of conventional security approaches, providing robust protection against future quantum computing capabilities.

The observed threat detection performance of QShield-ZTN is primarily driven by the coordinated interaction between its quantum-resilient security layer and autonomous orchestration mechanisms. Rather than relying solely on static classification, the framework continuously adapts detection behavior through dynamic trust enforcement and predictive orchestration, which improves discrimination between benign anomalies and true attack patterns. The strong performance against quantum-enhanced attacks can be attributed to the hybrid integration of post-quantum cryptography and quantum-aware threat modeling, which expands the effective security feature space and reduces ambiguity in adversarial scenarios. Moreover, the low false positive and false negative rates indicate that the zero-trust policy engine and multi-agent coordination stabilize decision boundaries over time, making the system suitable for fully autonomous operation where excessive alerts or missed detections would otherwise degrade network reliability.

Fig 4 presents a focused analysis of critical security metrics for different frameworks. Detection accuracy is shown to peak for QShield-ZTN at 97.3%, while error rate comparisons reveal lower false positive and negative rates for advanced models. The quantum resistance bar chart demonstrates QShield-ZTN and Post-Quantum Only models offer the strongest resistance (256 bits), positioning them as top candidates for future-proof security deployments.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Security performance comparison across three visualizations: (left image) detection accuracy line plot showing QShield-ZTN achieving 97.3% accuracy, (center image) error rates bar chart demonstrating lowest false positive (2.1%) and false negative (1.8%) rates, and (right image) quantum resistance horizontal bar chart displaying maximum 256-bit security strength compared to baseline methods.

https://doi.org/10.1371/journal.pone.0357590.g004

The post-quantum cryptographic components provide security levels equivalent to 256-bit classical security against quantum attacks, while maintaining efficient computation performance suitable for real-time 6G network operations. The hybrid cryptographic approach ensures continued security even in scenarios where individual cryptographic primitives may be compromised.

The zero-trust dynamic policy management achieves continuous verification with average policy update times of 15.3 milliseconds, enabling real-time adaptation to changing threat conditions without significant impact on network performance.

Table 6 presents the comprehensive security performance comparison, demonstrating QShield-ZTN’s superior performance across all security metrics. The results show significant improvements in detection accuracy, reduced false positive and negative rates, enhanced quantum resistance strength, and faster response times compared to existing methods.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Security Performance Comparison Results.

https://doi.org/10.1371/journal.pone.0357590.t006

Table 7 presents the per-dataset detection performance of QShield-ZTN across all four evaluation datasets. The results demonstrate consistent detection effectiveness regardless of dataset characteristics, with accuracy ranging from 96.8% on NSL-KDD+ to 97.8% on QShield-Synthetic. The highest performance on QShield-Synthetic is attributed to the richer feature space (156 features) that enables the ensemble detection model to better discriminate between benign and adversarial patterns. The slightly lower performance on NSL-KDD+ and Quantum-Dataset reflects their smaller feature sets (43 and 67 features, respectively), which constrain the discriminative capacity of the multi-modal detection pipeline. Importantly, the narrow confidence intervals across all datasets confirm the stability and reproducibility of the reported results.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Per-Dataset Detection Performance of QShield-ZTN.

https://doi.org/10.1371/journal.pone.0357590.t007

Fig 5 presents a comparative synthesis of security and orchestration performance across multiple frameworks using normalized metrics and component-level aggregation. The left subplot illustrates how detection accuracy, quantum resistance, response speed, and false positive behavior vary across methods, revealing distinct trade-offs between classical, hybrid, and quantum-aware approaches. Rather than excelling in a single dimension, higher-ranked frameworks demonstrate balanced performance across multiple metrics. The right subplot aggregates individual metric contributions into an overall performance ranking, highlighting how cryptographic strength, detection capability, orchestration efficiency, and resource utilization jointly influence system effectiveness. Frameworks relying primarily on classical or trust-based mechanisms exhibit uneven profiles, typically achieving reasonable detection performance but reduced resilience or slower response behavior. In contrast, hybrid and quantum-aware designs achieve more consistent performance across components.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Multi-metric performance comparison and overall ranking of evaluated security frameworks, illustrating normalized detection effectiveness, quantum resilience, response behavior, and component-level performance contributions.

https://doi.org/10.1371/journal.pone.0357590.g005

4.2. Network Performance Analysis

QShield-ZTN's performance gains stem from cross-layer co-design that jointly optimizes security enforcement and orchestration rather than treating them independently. Throughput improvements arise from proactive slice-level scheduling and load-aware resource allocation that anticipate congestion before it occurs, while latency reductions reflect decentralized agent-level decision-making that applies policy updates and routing adjustments ahead of threat or congestion propagation. Consistently low packet loss rates confirm that adaptive load balancing and intelligent routing maintain data-plane stability under simultaneous traffic and security stress, and energy efficiency gains emerge from coordinating cryptographic intensity, routing, and processing workloads with real-time network conditions rather than isolated hardware optimizations.

These performance metrics reflect real-time orchestration and control-plane decision latency under established secure sessions, excluding offline PQC setup and key-refresh overheads. Gains become more pronounced under higher network stress — as shown in the throughput-latency plot — where QShield-ZTN maintains stable operation as load increases, demonstrating that predictive orchestration and adaptive resource management sustain service quality under dynamic 6G conditions.

Table 8 compares the network performance of QShield-ZTN with existing frameworks in terms of throughput, end-to-end latency, packet loss rate, and energy efficiency. Results are reported as mean values with confidence intervals, showing that QShield-ZTN consistently achieves higher throughput, lower latency, reduced packet loss, and better energy efficiency under identical network conditions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Network Performance Metrics Comparison.

https://doi.org/10.1371/journal.pone.0357590.t008

4.3. Orchestration Efficiency Analysis

QShield-ZTN's performance gains stem from cross-layer co-design that jointly optimizes security and orchestration through proactive scheduling, decentralized decision-making, and coordinated power-aware resource allocation — maintaining throughput, low latency, and packet stability under simultaneous traffic and security stress. All reported latency values reflect real-time control-plane decisions under established secure sessions, with gains becoming more pronounced under higher network load, validating the effectiveness of predictive orchestration in dynamic 6G environments.

Fig 6 highlights the superior orchestration capabilities of QShield-ZTN versus state-of-the-art baselines. The plots show QShield-ZTN achieves consistently lower response times, fastest autonomous decision speed (8.7 ms), highest resource utilization efficiency (94.3%), and maintains robust scalability as network size increases up to 100,000 devices (scale factor 0.91). These results confirm QShield-ZTN’s advantage in efficient, scalable, and resilient autonomous orchestration for large-scale secure networks.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Real-time decision making and network scalability performance showing (top) QShield-ZTN maintaining consistent 8.9ms response time with 49.6% improvement over traditional SDN during network stress events, and (bottom) superior scalability maintaining 91% performance factor and 92% resource efficiency at 100K devices compared to competing methods.

https://doi.org/10.1371/journal.pone.0357590.g006

4.4. Quantum Resilience Evaluation

The quantum resilience evaluation assesses QShield-ZTN’s effectiveness against quantum computing threats and its ability to maintain security in post-quantum environments. The evaluation includes analysis of cryptographic strength, key distribution efficiency, and quantum attack resistance.

The post-quantum cryptographic strength analysis demonstrates that QShield-ZTN maintains security levels equivalent to 256-bit classical security against both classical and quantum attacks. The hybrid cryptographic approach provides redundant protection mechanisms that ensure continued security even if individual cryptographic primitives are compromised.

The quantum key distribution efficiency shows that QShield-ZTN achieves key generation rates of 1.2 MHz with 99.7% success rate under realistic quantum channel conditions. The adaptive quantum parameter optimization maintains high key generation rates even in the presence of quantum channel noise and eavesdropping attempts.

The quantum attack resistance evaluation demonstrates that QShield-ZTN successfully defends against various quantum attack scenarios including quantum cryptanalysis, quantum side-channel attacks, and quantum denial-of-service attacks. The multi-layered quantum-resilient architecture provides comprehensive protection across all network layers.

The quantum threat adaptation capability shows that QShield-ZTN dynamically adjusts security parameters based on quantum threat intelligence, maintaining optimal security-performance trade-offs as quantum computing capabilities evolve.

Table 9 presents the quantum resilience performance metrics, demonstrating QShield-ZTN’s superior quantum-resistant capabilities compared to existing methods. The quantum key distribution stream field illustrates how key generation efficiency is maximized by aligning channel conditions and orchestration decisions, reinforcing that quantum security effectiveness depends on dynamic coordination rather than fixed provisioning. The attack resistance surface further emphasizes this behavior by showing stable resistance across increasing attack complexity and time, indicating robustness against sustained adversarial pressure.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Quantum Resilience Performance Metrics.

https://doi.org/10.1371/journal.pone.0357590.t009

4.5. Comparative Performance Summary

The comparative evaluation indicates that the performance advantages of QShield-ZTN are not confined to a single metric but emerge from consistent improvements across security, orchestration, and network efficiency dimensions. Rather than optimizing one objective at the expense of others, the framework maintains balanced gains by jointly considering security strength, decision latency, scalability, and energy consumption during orchestration.

The most pronounced improvements are observed in quantum resilience and autonomous decision efficiency, which can be attributed to the architectural coupling of post-quantum cryptography with multi-agent learning-based control. By expanding the effective security space and enabling distributed decision-making, the framework reduces the need for conservative safety margins that often degrade performance in traditional designs.

From a cost–benefit perspective, the additional computational overhead introduced by quantum-resilient mechanisms is offset by reductions in retransmissions, policy re-evaluations, and reactive mitigation actions. This leads to net efficiency gains, particularly in large-scale deployments where orchestration overhead typically grows nonlinearly with network size.

The deployment readiness of QShield-ZTN is supported by its standards-aligned design and modular integration strategy. Rather than assuming fully quantum-enabled infrastructure, the framework supports phased adoption through hybrid cryptographic operation and backward-compatible orchestration interfaces, making it suitable for pilot deployment and incremental scaling as 6G infrastructure matures.

Table 10 summarizes the aggregate performance gains of QShield-ZTN across all major evaluation dimensions. The results indicate consistent improvements rather than isolated metric optimization, with the largest relative gains observed in quantum resilience and network performance. Improvements in orchestration efficiency and scalability are moderate but stable, reflecting the framework’s design trade-off between decision accuracy and operational overhead. The overall average improvement confirms that QShield-ZTN achieves balanced performance benefits across heterogeneous objectives without disproportionately favoring a single metric.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Overall Performance Improvement Summary.

https://doi.org/10.1371/journal.pone.0357590.t010

Table 11 extends the comparative analysis by including recently published named frameworks from 2024–2025 that address aspects of 6G security, zero-touch networking, or quantum-resilient communications. Unlike the category-based baselines in Tables 3–5–5, this comparison maps specific published approaches to the same evaluation dimensions used for QShield-ZTN. The results confirm that no existing single framework simultaneously addresses quantum-resilient cryptography, dynamic zero-trust enforcement, and autonomous AI-driven orchestration. SecOrch-6G [2] provides orchestration capabilities but lacks quantum resilience entirely. ZT-CPS-6G [3] implements zero-trust strategies but does not support adaptive orchestration or quantum-safe mechanisms. QT-B5G [4] incorporates QKD but omits PQC integration and autonomous decision-making. Among the most recent AI-focused approaches, Agentic-AI-Comm [17] and FL-LLM-Wireless [16] demonstrate autonomous capabilities but lack security frameworks and quantum resistance. QShield-ZTN uniquely integrates all three pillars—quantum resilience, zero-trust enforcement, and autonomous orchestration—within a unified, standards-compliant architecture, which accounts for the observed performance advantages across all evaluation dimensions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Extended Comparison with Recent Named 6G ZTN and Secure Orchestration Frameworks (2024–2025).

https://doi.org/10.1371/journal.pone.0357590.t011

The layer-wise security framework proposed by Yang et al. [40] provides an in-depth analysis of security vulnerabilities across the quantum internet stack, emphasizing cryptographic primitives, protocol-level defenses, and threat surfaces at different layers. This work offers valuable insights into where quantum-specific protections are required and how security guarantees propagate across layers.

QShield-ZTN complements this perspective by addressing how such layer-wise protections can be orchestrated autonomously at scale within 6G zero-touch networks. While [40] focuses on security analysis and architectural layering, QShield-ZTN emphasizes operationalization through AI-driven orchestration, dynamic trust management, and performance-aware security adaptation. The two approaches are therefore synergistic: the security models and threat analyses in [40] can inform policy design within QShield-ZTN, while QShield-ZTN provides a practical control framework to enforce and adapt these protections in real-time deployments.

Overall, QShield-ZTN differs from existing quantum networking and security studies by unifying quantum-resilient mechanisms, autonomous orchestration, and zero-touch operation within a single standards-aligned framework, bridging the gap between theoretical quantum security models and deployable 6G network management solutions.

To ensure reproducibility and mitigate concerns of result inflation, all reported improvements were validated through repeated simulations and cross-validation across heterogeneous datasets. The observed gains are consistently linked to two architectural factors: the enlargement of the security decision space through quantum-resilient primitives, and the reduction of orchestration latency through predictive, multi-agent policy optimization.

Fig 7 summarizes the learning behavior and stability characteristics of the evaluated orchestration frameworks. In subplot (a), QShield-ZTN converges more rapidly to a high normalized security reward, indicating efficient policy optimization and stable coordination among agents during training. Baseline methods exhibit slower convergence and lower asymptotic reward levels, reflecting less effective adaptation under dynamic security conditions. Subplot (b) highlights training stability through privacy variance error analysis. QShield-ZTN maintains consistently lower variance throughout training, demonstrating improved robustness against fluctuations introduced by distributed learning and privacy constraints. This stability is achieved without sacrificing convergence speed, indicating that the proposed orchestration strategy balances learning efficiency and privacy preservation more effectively than comparative approaches.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Multi-agent learning performance comparison illustrating (a) normalized security reward convergence across training phases and (b) privacy variance error evolution during training.

QShield-ZTN achieves faster convergence and improved training stability relative to baseline frameworks, demonstrating efficient learning dynamics and robust privacy-preserving behavior.

https://doi.org/10.1371/journal.pone.0357590.g007

Together, the two subplots provide a consolidated view of convergence efficiency and stability, eliminating the need for redundant metric-specific figures while capturing the core learning advantages of QShield-ZTN.

Table 12 presents a quantitative convergence analysis comparing QShield-ZTN against RL-capable baselines. QShield-ZTN converges in 142 ± 8 episodes, approximately 49.5% faster than Federated Learning and 19.3% faster than Hybrid Classical-Q. The lower training stability variance (σ = 0.012) confirms that the multi-agent PPO-based orchestration engine achieves robust convergence without oscillatory behavior. The wall-clock training time of 4.7 hours reflects the computational efficiency of the CTDE paradigm, where centralized training leverages global state information to accelerate policy learning while decentralized execution maintains scalability.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Training Convergence Analysis of QShield-ZTN vs. Baselines.

https://doi.org/10.1371/journal.pone.0357590.t012

Fairness and Neutrality of Experimental Comparisons

All baselines were re-implemented under identical preprocessing, dataset partitions, and hardware configurations with fairly tuned hyperparameters, ensuring no artificial constraints skewed comparisons. Reported gains represent averages across heterogeneous scenarios — margins narrow under low-load or benign conditions and widen under high dynamism and adversarial stress, reflecting QShield-ZTN's design as a robustness-oriented framework rather than a benchmark-specialized optimizer.

4.6. Ablation and Component Contribution Analysis

To better understand the individual contribution of each architectural component within QShield-ZTN, we conducted a controlled ablation study in which key modules were selectively disabled while keeping all other system parameters unchanged. The objective of this analysis is not to re-optimize the framework under reduced configurations, but to quantify the marginal performance degradation caused by the removal of specific components. This approach allows a clearer attribution of observed gains to the quantum security stack, zero-trust enforcement, and autonomous learning mechanisms.

Four ablated configurations were evaluated: (i) QShield-ZTN without Quantum Key Distribution (QKD), (ii) without post-quantum cryptography (PQC), (iii) without Zero-Trust Architecture (ZTA), and (iv) without reinforcement learning–based orchestration (RL). Each configuration was tested under identical traffic loads, threat scenarios, and dataset splits to ensure fairness.

Interpretation of Ablation Results

Table 13 results reveal that different modules dominate different performance dimensions. The removal of QKD and PQC primarily affects quantum resistance, confirming that cryptographic resilience is jointly provided by both components rather than by a single security primitive. In contrast, disabling ZTA leads to a noticeable drop in detection accuracy, indicating that continuous trust verification plays a central role in stabilizing security decisions under heterogeneous attack conditions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. Ablation Study – Impact of Removing Individual QShield-ZTN Modules.

https://doi.org/10.1371/journal.pone.0357590.t013

The most pronounced degradation in latency and energy efficiency occurs when reinforcement learning–based orchestration is removed. This observation confirms that the autonomous orchestration engine is the dominant contributor to real-time performance and resource efficiency, as it enables predictive control and adaptive coordination across distributed agents. Without RL-driven decision making, the system reverts to reactive orchestration behavior, leading to delayed responses and increased overhead.

Overall, this ablation study demonstrates that QShield-ZTN’s performance gains do not stem from a single dominant module but from the synergistic interaction between quantum-resilient security mechanisms, zero-trust enforcement, and learning-driven orchestration. The results validate the necessity of an integrated design for achieving balanced improvements across security, latency, and energy efficiency in autonomous 6G zero-touch networks.

Table 14 evaluates the cross-dataset generalization capability of QShield-ZTN by training on one dataset and testing on another without retraining. The maximum performance degradation is 8.5% (NSL-KDD+ → QShield-Synthetic), while the minimum degradation is 3.7% (QShield-Synthetic → CICIDS2017+). The consistently moderate drop across all cross-dataset pairs indicates that the learned detection and orchestration policies generalize across heterogeneous traffic distributions and attack profiles. The strongest generalization from QShield-Synthetic reflects the benefit of its larger feature space (156 features) and diverse attack modeling, which enables learning of more transferable representations. These results support the practical deployability of QShield-ZTN in real-world environments where traffic characteristics may differ from training conditions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. Cross-Dataset Generalization Results.

https://doi.org/10.1371/journal.pone.0357590.t014

4.7. Advanced Performance Characteristics and Standards Compliance Analysis

This subsection presents advanced performance metrics including energy efficiency, scalability, threat detection accuracy, and standards compliance for the QShield-ZTN framework across diverse operational scenarios. QShield-ZTN achieves 3.2 × J/bit energy consumption, representing 23.8% improvement over conventional approaches through optimized quantum-resilient cryptographic implementations and intelligent power management.

Table 15 demonstrates QShield-ZTN’s superior energy performance across all network components with significant power savings in adaptive power management modes. The threat detection accuracy evaluation gives full evaluation of the security capability of QShield-ZTN in different attack scenarios and network conditions. The framework has the best detection capabilities against classical and quantum improved attacks with the help of its advanced ensemble machine learning approach and quantum-resistant threat identification mechanisms.

Download:

• PNG
  larger image
• TIFF
  original image

Table 15. Energy Efficiency Performance Comparison.

https://doi.org/10.1371/journal.pone.0357590.t015

The threat detection analysis includes evaluation against several categories of threats including network-level attacks including distributed denial of service and man-in-the-middle attacks; application-level threats including injection of malware and attempts at data exfiltration; quantum-enhanced attacks using quantum computing capabilities for cryptographic attacks; and advanced persistent threats which are sophisticated and involve multiple attack vectors and adaptive strategies. The results show that QShield-ZTN has an average threat detection accuracy of 97.3% with respect to any attack behavior and especially good performance against quantum enhanced attacks (98.7% accuracy) and advanced persistent threats (96% accuracy). The low false positive and false negative rates of 2.1% and 1.8% respectively are an indication of the reliability of the framework for autonomous operation scenarios in which human intervention is minimized.

The ensemble machine learning approach plays a major role in the enhanced performance of detection by integrating various detection algorithms such as statistical anomaly detection, behavioral analysis, signature-based detection, and quantum-enhanced pattern recognition. The adaptive learning mechanisms enable continuous improvement of detection accuracy based on observed attack patterns and network behaviors.

Table 16 reports the per-attack-category detection performance of QShield-ZTN. The framework demonstrates particularly strong performance against quantum-enhanced attacks, with quantum cryptanalysis detection achieving 98.7% accuracy. This is attributed to the hybrid integration of post-quantum cryptography and quantum-aware threat modeling, which expands the effective security feature space. Advanced persistent threats represent the most challenging category at 96.0% accuracy, owing to their multi-vector nature and adaptive evasion strategies. Nevertheless, this performance significantly exceeds the best baseline (91.3% for Hybrid Classical-Q), confirming the effectiveness of the ensemble machine learning approach combined with zero-trust continuous verification.

Download:

• PNG
  larger image
• TIFF
  original image

Table 16. Detection Accuracy by Attack Category.

https://doi.org/10.1371/journal.pone.0357590.t016

5.1. Security Architecture Advantages

As shown in Table 10 (Security Effectiveness) and Fig 5, QShield-ZTN demonstrates measurable improvements in detection accuracy and quantum resistance compared to baseline methods, particularly under high-threat and multi-domain attack scenarios. These gains are most pronounced when quantum-enhanced attacks are present, where detection accuracy reaches 97.3 ± 0.6%, exceeding the strongest baseline by 2.9%. Under low-threat conditions, several baselines achieve comparable accuracy, indicating that QShield-ZTN’s advantage is scenario-dependent rather than uniform.

The combination of post-quantum cryptographic algorithms and quantum key distribution provides redundant security mechanisms that ensure ongoing security in the presence of potential security breaches in individual cryptographic primitives due to the fast development of quantum computing. This hybrid approach is a far step beyond single-layer security solutions based on single cryptographic mechanisms.

The dynamic implementation of the zero-trust architecture enables continuous verification and the adaptive enforcement of the policy based on the changing threat conditions and network environments. Unlike static security policies employed by conventional systems, QShield's adaptive approach ensures the optimal security-performance trade-offs under different operational conditions. The quantum-resilient threat detection mechanisms make use of a very advanced machine learning algorithms with quantum-resilient features to detect both the classical and quantum-augmented attack. The ensemble approach that combination of statistical, machine learning and quantum enhanced detection methods provides better accuracy than individual detection methods.

The autonomous security orchestration capabilities are crucial for enabling real-time response to security threats without requiring human intervention, which is important for 6G networks, which will need to operate with minimal human oversight. The smart decision-making algorithms ensure the right response measures are taken without affecting the network performance and user experience much.

The benefits of autonomous orchestration are quantitatively supported by the orchestration decision time reported in in Fig 7(a). QShield-ZTN achieves a mean orchestration latency of 8.7 ± 0.4 ms, representing a 29.8% reduction compared to the best baseline. However, under small-scale or low-load conditions (≤10K devices), federated learning and adaptive policy baselines demonstrate comparable responsiveness, indicating that the orchestration advantage becomes more evident at larger scales.

5.2. Standards Compliance and Interoperability

QShield-ZTN is designed for comprehensive standards compliance across leading standardization bodies to ensure seamless integration with emerging 6G architectures. The framework aligns with 3GPP Release 18 and beyond, supporting network slicing, edge computing integration, and AI-native operations, while meeting ITU-R IMT-2030 performance targets for throughput, latency, reliability, and energy efficiency. Security design incorporates IEEE 1363.1 quantum-safe standards, IEEE 802.1X zero-trust extensions, NIST post-quantum cryptography recommendations, and ETSI QKD specifications, ensuring long-term interoperability with both existing and future quantum-resilient network infrastructure.

Table 17 provides a feature-level qualitative comparison between QShield-ZTN and seven recently published frameworks. The comparison reveals that existing approaches typically address one or two dimensions of the 6G security-orchestration challenge but lack comprehensive coverage. QShield-ZTN is the only framework that integrates all evaluated features within a unified architecture, which explains the consistent quantitative advantages reported in the preceding tables. This holistic integration is the primary architectural differentiator, enabling cross-layer optimization that isolated approaches cannot achieve.

Download:

• PNG
  larger image
• TIFF
  original image

Table 17. Feature-Level Qualitative Comparison with Recent 6G ZTN Frameworks.

https://doi.org/10.1371/journal.pone.0357590.t017

5.3. Statistical Significance Analysis

To assure the robustness of the performance comparisons made between QShield-ZTN and baseline frameworks, we conducted a detailed statistical significance analysis. A one-way ANOVA test was performed across four important evaluation metrics: detection accuracy, orchestration latency, quantum resistance and energy-efficiency. The results indicate that there are statistically significant differences in the performance of QShield-ZTN with p-values <0.01 in all the datasets. Furthermore, 95% confidence intervals were calculated for each reported mean and this indicated stable improvements for minimal variance. Additionally, we used a paired t-test of QShield-ZTN against the best performing baseline to ensure that these improvements weren't explained by random improvements (p < 0.005). This statistical validation is utilized to validate the consistency, reliability and reproducibility of the gains reported by QShield-ZTN.

Table 18 presents the complete statistical significance analysis comparing QShield-ZTN against the best-performing baseline (Hybrid Classical-Q) across all primary evaluation metrics. All paired t-tests yield p-values below 0.001, confirming that the observed improvements are statistically significant and not attributable to random variation. Cohen's d effect sizes exceed 1.0 for all metrics, indicating large practical significance. The 95% confidence intervals for the mean differences confirm that QShield-ZTN's advantages are consistent and stable across repeated experimental runs. The one-way ANOVA F-statistics further confirm significant inter-method performance differences across all 11 evaluated frameworks. These results collectively validate that QShield-ZTN's improvements are both statistically reliable and practically meaningful.

Download:

• PNG
  larger image
• TIFF
  original image

Table 18. Detailed Statistical Significance Analysis (QShield-ZTN vs. Best Baseline).

https://doi.org/10.1371/journal.pone.0357590.t018

To ensure robustness of the statistical conclusions without relying on normality assumptions, Table 19 reports non-parametric Wilcoxon signed-rank test results for all evaluation metrics. The W-statistics and corresponding p-values confirm that QShield-ZTN's performance advantages remain statistically significant under distribution-free testing. The rank-biserial correlation coefficients (≥0.90 for all metrics) indicate near-perfect directional consistency, meaning QShield-ZTN outperforms the best baseline in virtually every paired comparison across all independent runs.

Download:

• PNG
  larger image
• TIFF
  original image

Table 19. Non-Parametric Validation (Wilcoxon Signed-Rank Test).

https://doi.org/10.1371/journal.pone.0357590.t019

Table 20 reports per-dataset one-way ANOVA results for detection accuracy across all 11 evaluated frameworks. The large eta-squared (η²) values (0.88–0.93) indicate that framework selection accounts for 88–93% of the observed performance variance, confirming that the choice of orchestration and security architecture has a dominant effect on detection outcomes. Post-hoc Tukey HSD tests confirm that QShield-ZTN significantly outperforms the next-best framework on every dataset (p < 0.001 in all cases).

Download:

• PNG
  larger image
• TIFF
  original image

Table 20. Per-Dataset One-Way ANOVA Results (Detection Accuracy Across All 11 Frameworks).

https://doi.org/10.1371/journal.pone.0357590.t020

5.4. Limitaiosn and Future Research Directions

Simulation-to-Reality Gap Evaluation in NS-3.39/QKDNetSim lacks real-world impairments. Physical layer: mmWave models miss environmental variability (weather, obstacles)—empirical studies show 15–30% higher path loss variance. Hardware: assumes ideal synchronization/error-free operation; real devices exhibit jitter, thermal noise. Scale: 100 nodes over 1800s; production may have 10K+ devices operating months—emergent behaviors unexplored. Mitigation: Validation on physical testbeds planned.

Computational Overhead PQC latency: Kyber-1024 ~ 0.8ms/operation; 10K connections/s requires 8s CPU time/s—infeasible without acceleration. QShield-ZTN performs PQC async during session init (assumes > 30s sessions); ultra-short connections problematic. QKD infrastructure: 50 km fiber, 1.2MHz rate; > 100km degrades exponentially; $500K + /link cost barrier. Energy: 5-15W/node; network-wide deployment = kilowatt-range. Solutions: PQC ASICs, lightweight variants needed.

Scalability Bottlenecks Training: 142 episodes (10 agents); 50 agents → 3.2 × slower (super-linear). Centralized aggregation: single point of failure; 100 agents reporting simultaneously overwhelms coordinator. State space: O(N²) pairwise features; > 500 devices → 250K dimensions, challenging NN capacity. Solutions: Hierarchical orchestration, state abstraction, decentralized consensus.

Adversarial Robustness Threat model assumes static attacks. Adversarial ML: FGSM attacks degrade accuracy 12–18%. Policy evasion: attackers may reverse-engineer trust function. Quantum attacks beyond cryptanalysis unexplored. Mitigations: adversarial training, policy obfuscation needed.

Interoperability Gaps: Standards-compliant design untested in practice. Protocol compatibility with 5GC unvalidated. Vendor heterogeneity (Ericsson/Nokia/Huawei) not tested. 6G standards evolving (3GPP Rel 20 + targeting 2028)—design may require revision. Roadmap: standardization participation, vendor testing.

Deployment Challenges: Legacy coexistence strategies absent. Operational complexity requires expert configuration. Regulatory restrictions on QKD possible. Cost-benefit analysis unclear. Solutions: deployment guides, automated tuning tools needed.

Failure Scenarios: RL policy crash → loss of autonomy; no fallback. PQC vulnerabilities → security collapse; formal verification absent. QRNG/HSM failures → entropy/key management compromise; no redundancy. Cascading failures across slices unexplored. Future: fault-tolerance analysis, graceful degradation design.

Future Directions: Real-world testbed validation, scalability enhancements (hierarchical orchestration, state abstraction), adversarial hardening, lightweight PQC for IoT, interoperability testing, economic analysis, formal verification.

This transparency acknowledges QShield-ZTN as a research prototype demonstrating feasibility, not a production-ready system.