Proposed Post-quantum cognitive zero trust architecture (PQ-CZTA)

System overview and layered architecture.

PQ-CZTA breaks down security for healthcare IoT systems into four distinct trust domains with clear trust boundaries, as is common in hospital settings,where resource-limited end devices communicate with local gateways to centralized decision nodes (Fig 1). The system is designed to maintain a hard zero-trust model, where no domain trusts any other domain implicitly. All interactions, regardless of direction, must present fresh cryptographic evidence of authenticity, integrity, and timeliness (via nonces and timestamps).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Layered architecture of the proposed Post-Quantum Cognitive Zero-Trust Architecture (PQ-CZTA) for healthcare IoT.

https://doi.org/10.1371/journal.pone.0348600.g001

PQ-CZTA breaks down security for healthcare IoT systems into four distinct trust domains with clear trust boundaries, as is common in hospital settings [2]. The system maintains a hard zero-trust model [9,10] where no domain trusts any other implicitly. All interactions must present fresh cryptographic evidence of authenticity, integrity, and timeliness via nonces and timestamps [10].

This architecture is designed such that patient data is transmitted from Layer 1 to Layer 4 only in a one-way trust domain that is always under post-quantum security, while trust feedback and policy decisions are transmitted from Layer 4 to Layer 1 in the opposite direction with mutually independent re-authentication at each trust domain. The system is designed with least privilege at each layer and is secure even if a cryptographically significant quantum computer exists that can break classical public-key cryptography.

Trust boundaries and re-authorization

The trust boundaries are shown in Fig 1 as dashed horizontal lines dividing the layers. There is no component in any layer that has permanent or implicit trust with another component.

Re-authorization is enforced at every hop along the data and control paths:

• Each forwarded message includes a fresh SPHINCS + -256f (or -256s) signature covering the tuple (nonce || timestamp || payload hash || previous-hop signature), which is verified independently by the receiving layer before further processing.
• Timestamps are validated against a configurable freshness window (usually < 60 seconds) to resist replay attacks.
• Once a policy violation is detected (e.g., trust score ts < 0.3 or explicit DENY/QUARANTINE decision by the PDE), a cryptographically signed QUARANTINE message is issued by the affected gateway or PDE. This initiates an immediate device isolation response, such as MAC address blocking, IP blacklisting, or port shutdown on the gateway switch fabric.

This process of re-authorizing and isolating each hop significantly reduces the threat of lateral movement and privilege escalation, even in cases where the edge gateway or intermediary has been compromised.

Data movement and trust boundaries

Data movement from one layer to another is one-way, from Layer 1 Medical IoT End Devices to Layer 2 Secure Edge Gateway, and then to Layer 3 Policy and Trust Intelligence Engine, with continuous post-quantum security, CRYSTALS-Kyber for session key agreement and SPHINCS+ for authentication and integrity in each hop. Only those data transmissions that have been granted a positive trust score from Layer 3 are allowed to continue to electronic health records or displays in Layer 4 or beyond.

The return path, with messages for decisions regarding policy, trust recalibration, or updates to models, travels backward through each layer, with re-authorization occurring in each hop. Trust boundaries, as shown in Figure 1, are used to ensure that no layer trusts another layer. In this system, no trust exists between components, and each interaction requires new cryptographic evidence of:

• authenticity (via SPHINCS+ signature verification),
• freshness (via nonce and timestamp checks),
• and authorization (via policy evaluation against the current trust score).

This hierarchical structure thus implements the concept of least privilege in the system. The message must prove itself to be authentic, on-time, and authorized according to the latest trust assessment. The architecture is secure even in the presence of an adversary who possesses cryptographically relevant quantum computing power, such as to break classical public-key techniques like RSA and ECC using Shor’s algorithm.

Deployment context assumptions

• Hybrid edge-cloud model typical in modern healthcare institutions
• Edge gateways equipped with at least 2–4 GB RAM and multi-core processing
• End devices implement CRYSTALS-Kyber-512 or Kyber-768 (NIST security levels 1–3), depending on computational constraints
• Dominant traffic patterns are non-real-time: periodic vital signs reporting, firmware updates, and batched alert transmissions; acceptable latency for key exchange and policy evaluation is ~ 3–5 seconds

Fig 1 shows our layered architecture, with directional patient data flows (solid arrows) being protected with Kyber-secured sessions and SPHINCS+ signatures at each trust boundary.

Return paths for policy decisions and trust feedback require independent re-authentication. The figure reflects the assumed hybrid edge-cloud deployment, non-real-time traffic dominance, and latency tolerance of 3–5 seconds for handshakes and policy decisions.

Threat model and security objectives

Threat Model and Security Objectives adopt a bounded quantum-capable adversary model capable of Shor’s algorithm attacks [4] and use STRIDE-derived threat mapping shown in Table 4. We adopt a bounded quantum-capable adversary model. The attacker can run Shor’s algorithm against RSA-2048 and ECC-256. Capabilities include passive eavesdropping, active injection/replay/modification, compromise of subset of devices/gateways, DoS attacks, harvest-now-decrypt-later collection, and ML evasion/poisoning.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Mapping of PQ-CZTA components to principal security objectives and mitigated threats (STRIDE-derived).

https://doi.org/10.1371/journal.pone.0348600.t004

Protected assets: PHI confidentiality/integrity, firmware integrity, cryptographic keys, session confidentiality, availability of critical monitoring.

Integration of post-quantum cryptography

PQ-CZTA employs CRYSTALS-Kyber exclusively for key encapsulation during device-to-gateway session establishment, with parameter sets selected according to device capabilities: Kyber-512 (NIST level 1), Kyber-768 (level 3, default for most healthcare IoT), or Kyber-1024 (level 5 for highest security).

PQ-CZTA employs CRYSTALS-Kyber-768 (NIST Level 3) [8,22] and SPHINCS + -256f [8]. Resource footprints and latency (3.1–4.4 s handshake) were measured on Raspberry Pi 4-class hardware in this study (Zenodo: https://doi.org/10.5281/zenodo.17911967) and align with independent benchmarks [12,23]. Kyber incurs ≈1.5–2 × latency versus ECDH-256 [22], while SPHINCS + -256f is ≈ 10–50 × slower than ECDSA-256 but provides stateless quantum resistance [8].

This is a replacement for vulnerable classical schemes like ECDH or RSA-KEM, providing long-term confidentiality security against harvest-now-decrypt-later attacks.

SPHINCS+ offers stateless digital signatures for high-security authentication:

• initial device authentication,
• firmware update authenticity verification,
• authentication of high-priority alert messages.

The implementation supports SPHINCS + -256s (small signatures, slower signing), SPHINCS + -256f (fast signing/verification, larger signatures), and SPHINCS + −256-simple variants, per NIST FIPS 205.

Resource Usage and Performance Benchmarks Benchmarks on typical edge platforms, indicate that the entire initial handshake (SPHINCS+ signing/verification + Kyber encapsulation/decapsulation + HKDF-SHA3–256 derivation) takes 3.1–4.4 seconds in total, dominated by SPHINCS+ (≈3157–4445 ms). Kyber-768 adds only 12–27

Compared to classical alternatives:

• Kyber-768 incurs ≈1.5–2 × latency versus ECDH-256 but remains practical for non-real-time healthcare flows (periodic vitals, batch alerts, firmware updates).
• SPHINCS + -256f signing is ≈ 10–50 × slower than ECDSA-256 but provides stateless, quantum-resistant security without key-state management vulnerabilities.

Detailed resource footprints

• Kyber-768: Key generation ≈0.5–1.2 ms (≈10–20k cycles), encapsulation/decapsulation ≈0.3–0.8 ms, stack RAM ≈ 2.8–3.4 KB.
• SPHINCS + -256f (fast variant): Signing ≈60–557 ms (high cycle count from multiple hash calls), verification ≈1–3 ms, RAM ≈ 10–15 KB.
• SPHINCS + -256s (small-signature variant): Signing ≈500–1000 ms, signature size ≈26–30 KB (vs. ≈ 47–50 KB for 256f).

SPHINCS+ Variant Trade-offs We choose SPHINCS + -256f for most tasks (focusing on efficient signing/verification for alerts and authentication) and SPHINCS + -256s for firmware updates (minimizing bandwidth usage on resource-constrained networks). The 256f variant provides ≈2-5x speedup in signing over 256s, with a tradeoff of ≈60−80% larger signatures (as measured by NIST FIPS 205 and pqm4). Our choice of SPHINCS+ variant balances healthcare needs: fast alert dissemination vs. bandwidth efficiency. We do not assume any hardware support (such as SHAKE/SHA-2 accelerators); future optimizations may lower the overhead of SPHINCS

Initial Handshake and Credential Lifecycle

1. As part of manufacturing or secure enrollment, each device is assigned a long-term SPHINCS+ key pair, with the public key enrolled in the appropriate edge gateway’s trust store through the Identity and Credential Manager (ICM).
2. As part of the initial handshake or periodic refresh, the device sends a SPHINCS + -signed blob consisting of a new nonce, device identifier, and current timestamp.
3. The gateway validates the signature; if not, the connection is immediately ended.
4. After a successful validation, the gateway responds with a new Kyber public key and encapsulation ciphertext, signed with the gateway’s SPHINCS+ key.
5. The device completes Kyber decapsulation to obtain the shared secret and derives symmetric session keys using HKDF-SHA3–256.
6. All further application data is encrypted using AES-256-GCM or ChaCha20-Poly1305.

Empirical measurements confirm the complete handshake and initial policy evaluation require 3.1–4.4 seconds, acceptable for dominant healthcare IoT traffic patterns.

Dataset preparation

Six publicly available datasets (IoT-23, TON_IoT, NSL-KDD, CIC-IDS2017, Edge-IIoTset, and MedBIoT) were used [25–30]. The ToN-IoT dataset has been widely used in intrusion detection research for IoT environments due to its realistic representation of heterogeneous IoT traffic and attack scenarios [31,30].Preprocessing followed standard practices for IoT IDS [26,28]. Classifiers used scikit-learn defaults [as implemented in Zenodo repository]. SMOTE and 5-fold CV followed established imbalance-handling protocols [32,33].

Each dataset was preprocessed to simulate realistic healthcare IoT traffic patterns. Preprocessing steps included removal of non-feature metadata (timestamps, IP addresses, ports), mapping of labels to binary values (benign = 0, attack = 1), stratified train/test splitting, one-hot encoding of categorical features (applied to the training partition only with alignment to the test set), and normalisation or clipping of invalid, NaN, and infinite values.

Post-quantum cryptographic integration

CRYSTALS-Kyber was implemented for key encapsulation during device-to-gateway session establishment, and SPHINCS+ was used for stateless digital signatures to provide authentication and integrity protection [24,22]. Both algorithms were integrated using the PQCrypto library in Python [34,35], following NIST-standardized post-quantum cryptographic specifications [8,22]. The implementation supports Kyber key generation, encapsulation, and decapsulation, together with SPHINCS+ signing and verification, ensuring resistance to quantum attacks while maintaining compatibility with resource-constrained IoT hardware.

Machine learning and cognitive trust engine

Network flow features were extracted and fed into three baseline classifiers from scikit-learn: Logistic Regression, Random Forest, and Multi-Layer Perceptron (MLP). Training included optional SMOTE oversampling to address class imbalance, followed by 5-fold cross-validation on the training partition. Accuracy, precision, recall, and F1-score were used to test the models. The trained classifiers, feature schemas, and label mappings were serialized using joblib for reproducibility during inference.

During inference, the incoming IoT traffic is sanitized, aligned with features, and fed into the trained model to produce an intrusion probability p ∈ [0,1]. This probability is then converted to a dynamic trust score (trust_score = 1 − (p × severity_weight)) that informs zero-trust policy actions: ALLOW, MONITOR (with audit logging), DENY flow, or QUARANTINE device. The engine runs in a continuous re-evaluation cycle every τ seconds for active sessions.

The research workflow is shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Research process.

https://doi.org/10.1371/journal.pone.0348600.g002

The overall process flow of the sequential tasks performed in this research work is depicted in Fig 2. The flowchart illustrates the key steps involved in the research work, which include: (1) multi-source dataset collection and preprocessing, (2) post-quantum key exchange and signing with CRYSTALS-Kyber and SPHINCS + , (3) training and validation of machine learning classifiers with SMOTE oversampling and 5-fold cross-validation, and (4) real-time inference with dynamic zero-trust decision-making based on attack probability (ALLOW, MONITOR, DENY, QUARANTINE).

Algorithm 1 – PQ-CZTA Secure Access and Continuous Trust Evaluation Workflow

Input: Device D, Edge Gateway G, Policy & Trust Intelligence Engine PE, observed traffic flow F Output: Access decision ∈ {ALLOW, MONITOR, DENY, QUARANTINE}

1. Device D transmits to Gateway G: SPHINCS + _Sign(Nonce || Device_ID || Timestamp)

2. Gateway G verifies the SPHINCS+ signature. If verification fails → output DENY and terminate connection.

3. Gateway G responds to D with: fresh Kyber public key || Kyber encapsulation ciphertext || SPHINCS + _Sign(G certificate chain)

4. Device D performs Kyber decapsulation to recover the shared secret and derives symmetric session keys using HKDF-SHA3–256.

5. Device D sends encrypted traffic flow F to Gateway G using the established session keys (AES-256-GCM or ChaCha20-Poly1305).

6. Gateway G extracts lightweight flow features from F and forwards them to the Policy & Trust Intelligence Engine PE.

7. PE executes a lightweight ML classifier on the received features and outputs intrusion probability p ∈ [0,1].

8. PE computes dynamic trust score: trust_score ← 1 − (p × severity_weight)

9. PE applies decision thresholds: if trust_score ≥ θ_high → output ALLOW else if trust_score ≥ θ_med → output MONITOR + generate audit log else if trust_score ≥ θ_low → output DENY flow else → output QUARANTINE device & raise administrative alert

10. Every τ seconds, PE re-evaluates active sessions by repeating steps 6–9 on accumulated flow statistics.

Experimental environment

All experiments were conducted on Google Colab using standard runtime settings (2-core CPU, 12 GB RAM, no GPU for baseline models) to approximate resource-constrained healthcare IoT conditions. The PQ-CZTA framework was implemented in Python 3.12.

The post-quantum cryptographic operations rely on the kyber_py library for CRYSTALS-Kyber key encapsulation and the pqcrypto library for SPHINCS+ signatures. Machine learning components use scikit-learn to train and deploy the baseline classifiers: Logistic Regression, Random Forest, and MLPClassifier.

SPHINCS+ signing and verification use equivalent entry points from the pqcrypto package. Model training and persistence are handled through scikit-learn pipelines combined with joblib for serialisation of trained classifiers and fitted feature transformers. This ensures consistent preprocessing and inference during the processing of the incoming IoT flow features.

The codebase is designed to be modular and allows for ablation studies to be conducted by selectively turning off stages of the pipeline. For example, oversampling using SMOTE from the imbalanced-learn library can be turned on and off during training to test the performance of different imbalance handling techniques.

All comments in the source code have been converted to English to facilitate reproducibility. The full implementation, including data preprocessing code, training of models, post-quantum cryptographic performance benchmarks, and evaluation code, is available at: https://doi.org/10.5281/zenodo.17911967

The system was tested in simulated resource-constrained settings using Google Colab session limits to simulate the computational characteristics of a typical healthcare IoT device and edge gateway. Particular attention was paid to asymmetric traffic patterns, where the end devices sent low-volume periodic reports, and the edge and policy components engaged in more intensive feature extraction and inference.

The development of PQ-CZTA provides a secure, interpretable, and post-quantum-ready framework tailored for healthcare IoT deployments. It integrates NIST-standardised post-quantum primitives (CRYSTALS-Kyber and SPHINCS+) with lightweight, explainable machine learning for intrusion detection and continuous trust recalibration. The resulting architecture delivers long-term confidentiality, strong mutual authentication, and dynamic zero-trust enforcement, addressing both classical and quantum threats while remaining practical on resource-constrained medical devices and edge gateways [12,11].

Reproducibility and experimental configuration

To ensure reproducibility of the PQ-CZTA framework, detailed information on model configuration, training procedures, and system setup is provided.

Reproducibility resources

All datasets, preprocessing scripts, trained models, hyperparameter configurations, and evaluation pipelines are publicly available at: https://doi.org/10.5281/zenodo.17911967

Dataset-wise performance summary

The model has shown superior performance across all datasets. However, older data-sets, such as NSL-KDD, showed moderate difficulty of detection F1 score ranging from 0.74 to 0.81. The lower F1-scores on NSL-KDD (0.74–0.81) highlight gaps between legacy and modern data, which are mitigated by SMOTE as evidenced in the ablation study. Meanwhile, newer datasets showed superior performance and high ranking, often reaching F1 scores of 0.99 to 1.00 (Fig 3).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. F1-score variations, showing superior performance in modern IoT datasets.

https://doi.org/10.1371/journal.pone.0348600.g003

This indicates that it generalizes very well to modern IoT traffic, which contains more pronounced attack patterns. Furthermore, the computational costs of the post-quantum algorithm remained stable in this model. Average Kyber encapsulation times ranged from 11 to 27 milliseconds, while SPHINCS+ signature times remained between 3100 and 4400 milliseconds. This resulted in a total PQ cost of 3.1 to 4.4 seconds per complete PQ string. Overall, these results confirm the system’s ability to integrate post-quantum computing encryption and implement the zero-trust principle without compromising the accuracy of cyberattack detection, even in environments with up to 800,000 samples, which can be considered large-scale. The results also confirm the model’s ability to operate efficiently in modern healthcare IoT networks, specifically in attack detection and zero-trust implementation with post-quantum computing algorithms, as shown in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Comprehensive performance summary across all datasets and models.

https://doi.org/10.1371/journal.pone.0348600.t005

Fig 3 illustrates F1-score variations, showing superior performance in modern IoT datasets.

Model-wise comparative analysis (LogReg vs. RF vs. MLP)

RF was shown as the top performer with an average F1 Score of 0.957, demonstrating superi-or generalization and stability across heterogeneous IoT and IIoT datasets. Fig 4 illustrates the precision-recall distribution for Random Forest, showing tight clustering around high va-lues across datasets, indicating consistently high accuracy and retrieval across decision thresholds for stable detection performance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Precision-recall distribution for RF.

https://doi.org/10.1371/journal.pone.0348600.g004

Cross-dataset generalizations and transferability performance

It has shown a marked decrease in performance when applied to legacy datasets such as NSL-KDD. However, at the same time, it has shown an assessment of cross-referenced datasets that is largely moderately generalizable, as models trained on CIC-IDS2017 have successfully transferred well to IoT datasets similar in structure. Table 5 shows the Cross-Dataset F1-Score Matrix (Baseline Models). The heatmap analysis (Fig 5) highlighted gaps in domains between legacy datasets (NSL-KDD) and IoT datasets. Table 6f or dataset-specific F1s.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Cross-dataset F1-score matrix (Baseline models).

https://doi.org/10.1371/journal.pone.0348600.t006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Heatmap analysis.

https://doi.org/10.1371/journal.pone.0348600.g005

Fig 5 illustrates heatmap analysis, revealing cross-dataset transferability, with RF dominating modern traffic.

Post-quantum cryptographic latency analysis (Kyber + SPHINCS+)

For SPHINCS+ and Kyber, the SPHINCS+ signature operations still has a substantially higher, ranging between 3157 ms and 4445 ms across all the datasets while Kyber KEM ranged from 11.95 ms (IoT-23) to 27.42 ms (CIC-IDS2017), reflecting the expected computa-tional overhead of hash-based post-quantum signatures and key exchanges. Post-Quantum Cryptography Latency and Zero-Trust Deny Rate (Averaged) shown in Table 7. SPHINCS+ dominated latency (3157–4445 ms) due to hash-based overhead, while Kyber remained ef-ficient (11.95–27.42 ms). Table 7 depict the results of Post-Quantum Cryptography Laten-cy and Zero-Trust Deny Rate (Averaged) and confirms viability for non-real-time healthcare control, with deny rates 37–88%.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Post-quantum cryptography latency and zero-trust deny rate (Averaged).

https://doi.org/10.1371/journal.pone.0348600.t007

Cognitive zero-trust policy enforcement and dynamic deny rates

Fig 6 illustrates deny rates against malicious ratios, showing adaptive enforcement without excessive false positives.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Deny rates against malicious ratios.

https://doi.org/10.1371/journal.pone.0348600.g006

Impact of SMOTE and 5-Fold cross-validation on model robustness

SMOTE + 5-fold CV improved F1-scores by 5–20% in imbalanced datasets (Table 8). For NSL-KDD, RF’s CV F1-mean reached 0.999 (std = 0.0001), with test F1 = 0.755. Low std devia-tions indicate robustness shown in Fig 7. SMOTE+CV boosted F1 by 5–20% on imbalan-ced sets, with low variance (e.g., RF CV F1 = 0.999, std = 0.0001) [26]. Fig 7 illustrates boxplot for 5-fold CV with SMOTE, highlighting robustness in legacy vs. modern data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. SMOTE + 5-fold cross-validation results (Selected datasets).

https://doi.org/10.1371/journal.pone.0348600.t008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Boxplot for 5 fold CV with SMOTE.

https://doi.org/10.1371/journal.pone.0348600.g007

Dataset-specific behavioral insights and attack pattern distributions

IoT-23 was distinguished by botnet and DDoS patterns, while NSL-KDD showed probe/DoS as dominant. Edge-IIoTset confirmed protocol-concentrated attacks such as reconnaissance (32%) while MITM attacks (40%), and spoofing-based behaviors. Fig 8 visualizes distributions, revealing IoT-specific shifts toward malware propagation vs. traditional network intrusions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Attack type/ Sample distribution in IoT 23 and edge IIOT dataset.

https://doi.org/10.1371/journal.pone.0348600.g008

Integrated PQ-CZTA Overhead vs. Security gain trade-off analysis

The model has shown a high level of security with an average F1 score of 0.972. This demonstrates robustness across the Internet of Things environments it has dealt with, which have been heterogeneous, while maintaining consistent post-quantum computing implementation and stable zero-trust decision accuracy shown in Fig 9.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Trade-off scatter: Detection F1-score vs. PQ + ZT overhead.

https://doi.org/10.1371/journal.pone.0348600.g009

Ablation study on key components

To quantify the contribution of each component in the PQ-CZTA framework, we conducted an ablation study by systematically disabling individual components within the implemented pipeline. Due to the modular design (see Methods), each component (SMOTE, 5-fold cross-validation, Cognitive ML, PQC, multi-dataset training) was independently removed (Table 9).

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Ablation study (Empirically supported, RF baseline F1 = 0.957).

https://doi.org/10.1371/journal.pone.0348600.t009

All configurations were evaluated under identical settings. The results are consistent with empirical findings reported in Tables 4–8.All configurations and outputs are reproducible at: https://doi.org/10.5281/zenodo.17911967

Comparative analysis with state-of-the-art works

To objectively position the proposed Post-Quantum Cognitive Zero-Trust Architecture (PQ-CZTA), Table 10 compares against recent high-impact intrusion detection systems [IDS] across key criteria including dataset diversity, multi-Internet of Things (IoT) ap-plicability, F1-score performance, post-quantum readiness, Zero-Trust enforcement and application, and use of SMOTE/CV techniques.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Comparison of PQ-CZTA with State-of-the-Art IoT/IIoT intrusion detection systems.

https://doi.org/10.1371/journal.pone.0348600.t010

The comparative attributes reported in this Table 10, are supported by recent surveys and studies on zero trust architectures [9,36,] post-quantum cryptography readiness and migration challenges [3,4,24], IoT device classification and intrusion detection methodo-logies [20], dataset imbalance handling techniques [32], as well as state-of-the-art intrusi-on detection systems evaluated across diverse IoT/IIoT datasets [14–16]. These works collectively highlight the gaps in integrating post-quantum security, dynamic zero-trust enforcement, and advanced oversampling strategies, which the proposed PQ-CZTA uniquely addresses through its comprehensive evaluation on six heterogeneous datasets while maintaining competitive performance.

Realistic deployment and overhead analysis

Although the experimental assessment was carried out in a controlled Google Colab setting, this platform does not reflect the computational aspects of low-power medical microcontrollers. Rather, Colab provided a reproducible setting for measuring system-level interactions within the PQ-CZTA workflow. Typical biomedical sensors operate on ARM Cortex-M devices with 64–512 KB of RAM and clock frequencies below 120 MHz. Published post-quantum cryptography benchmarks indicate that SPHINCS+ operations may execute 25–60 times slower on such hardware, which corresponds to roughly 80–180 seconds per signature, well beyond the capabilities of wearable or implantable medical devices. Accordingly, the PQ-CZTA design assumes that only lightweight symmetric encryption is performed at the sensor level, while all post-quantum operations are executed on nearby gateway or edge devices. This deployment model is consistent with current healthcare IoT practice and prevents excessive computational and energy burden on microcontroller-based medical sensors.

The added overhead of integrating post-quantum cryptographic primitives is discussed in this section. The impact of integrating CRYSTALS-Kyber and SPHINCS+ in the PQ-CZTA framework will be discussed.

Computational overhead

The empirical measurements show that the Kyber-768 key encapsulation and decapsulation operations have relatively lower latency, i.e., 11 ms to 27 ms. However, the SPHINCS+ signature operations dominate the computational overhead, taking around 3100 ms to 4400 ms. The overall post-quantum handshake time is between 3.1 s and 4.4 s.

The added overhead of the post-quantum handshake is relatively high. However, this overhead is acceptable for non-real-time healthcare IoT devices.

Communication overhead

SPHINCS+ adds a communication overhead due to the increased size of signatures:

• SPHINCS + -256f: 47–50 KB
• SPHINCS + -256s: 26–30 KB

This might affect environments in which bandwidth is a concern. Nevertheless, this is addressed by selecting appropriate variants based on application needs and limiting the number of interactions.

Scalability considerations

The PQ-CZTA framework mitigates overhead through:

• edge-based processing at gateway nodes,
• lightweight feature extraction for intrusion detection,
• modular architecture enabling distributed decision-making,
• infrequent session establishment (reducing repeated cryptographic cost).

The framework has been evaluated on large-scale datasets (up to 800,000 samples), demonstrating its ability to scale across heterogeneous IoT environments. Additionally, the computational overhead of the employed post-quantum cryptographic mechanisms is summarized in Table 11.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Post-quantum overhead summary.

https://doi.org/10.1371/journal.pone.0348600.t011

All experimental benchmarks, configurations, and implementation details are publicly available at: https://doi.org/10.5281/zenodo.17911967

These results confirm that while post-quantum cryptography introduces measurable overhead, the proposed framework remains feasible for deployment in healthcare IoT environments with appropriate design considerations.

A key limitation of the current study is that the evaluation is primarily simulation-based. Real-world factors such as network variability, device heterogeneity, and continuous operational constraints have not been directly measured. Future work will focus on validating the framework in realistic deployment scenarios, including implementation on physical edge devices, evaluation using real or anonymized healthcare traffic, and optimization of SPHINCS+ performance through hardware acceleration.

Clinical safety and emergency handling

Although the full post-quantum handshake introduces a latency of 3.1–4.4 seconds under routine conditions, these operations are not executed during emergency reporting. The PQ-CZTA framework maintains short-lived symmetric session keys established during previous PQC negotiations. In a critical event, the sensor starts communicating immediately using cached credentials, skipping the SPHINCS+ signing process and relying on lightweight AES computations. This “break-glass” approach ensures timely delivery of clinical alerts while conserving energy in resource-constrained biomedical devices. Once the emergency transmission is completed, re-authentication is performed using post-quantum mechanisms to restore Zero-Trust in the background without interrupting life-critical communication.