Evaluating Deep Belief Networks on IoMT ‎Dataset: The Impact of Different Network Designs on Attack Detection ‎Accuracy

The Internet of Things (IoT) is a fast-changing field allowing communication between electronic sensors and devices that are connected to the internet. One important use of this technology is Internet of Medical Things (IoMT), where a connection is made between healthcare services and patients by using many sensors, like electrocardiography devices, heart rate monitors, and blood pressure tools [1].

These developments have changed patient care in a significant way with remote diagnosis and health monitoring, which will continue for a long time. However, the wide connection and constant data exchange inside IoMT systems have created serious cybersecurity problems [2, 3]. Sensitive medical data becomes the main target for attackers, which can cause dangers with data change, man-in-the-middle attacks and denial of service attacks [4, 5]

Many medical devices have limited resources and sufficient computational power and energy are not available; also, traditional security methods like complex encryption or strong firewalls cannot be supported easily. Because of this situation, an urgent need is seen for advanced intrusion detection systems (IDS) that are specially designed for the healthcare environment [6, 7].

Artificial intelligence (AI) has an important role in improving security by detecting unusual behavior. Traditional machine learning (ML) methods have been used before in IDS research, but now the network traffic in IoT has become larger and more complex, so deeper models are needed. Deep learning (DL) methods, such as deep neural networks (DNN) and recurrent neural networks (RNN), are considered better solutions because they can learn by themselves and also understand complex patterns from high-dimensional data [6, 8, 9].

Among many DL methods, the Deep Belief Network (DBN) is seen as an important way for finding network attacks. A DBN is made from many layers of random hidden variables, and it is often arranged like a stack of Restricted Boltzmann Machines (RBMs). These structures are built to learn mixed and hierarchical data forms, and they are very useful for feature learning and also for reducing dimensionality in network intrusion detection situations [1, 6].

The performance of a DBN-based IDS is strongly affected by its design, including how many hidden layers exist and how many units are inside each layer. In present studies, three hidden layers are often used by default; however, there is still no clear agreement about the best number of neurons for getting high detection results [6, 10, 11].

This study aims to examine how different DBN architectures affect attack detection accuracy using publicly available benchmark datasets, namely CICIoMT2024 and Edith Cowan University-Internet of Health Things (ECU-IoHT), by investigating how architectural variations in depth, width, regularisation, and activation function influence detection performance against threats such as spoofing, Denial of Service (DoS), and reconnaissance.

The growth of the IoMT changed healthcare in a strong way, with real-time monitoring and data-based insight, many cybersecurity weaknesses have also been created. Because healthcare data is very sensitive and many medical devices have limited resources, traditional security methods like sophisticated firewalls are often not possible to apply, and therefore intelligent IDS systems become necessary [4, 10].

1.1 Traditional classifiers in Internet of Medical Things Security

ML methods have been widely studied for detecting attacks in IoMT network traffic. Algorithms like Support Vector Machines (SVMs), Random Forest, and K-Nearest Neighbors (KNN) show noticeable performance in many research areas. Random Forest is often seen as a very reliable method for stable threat detection and sometimes reaching nearly perfect accuracy in certain cases [12, 13].

However, an important limitation in many studies exists, since models are tested in closed-set conditions, which do not include new or unknown attack types or even small changes in known attacks. Also, shallow ML models have difficulty when dealing with high-dimensional data, which is very common in modern IoMT systems; therefore, more advanced architectures are needed [8, 14].

1.2 The Evolution toward Deep Learning and Deep Belief Networks

DL is becoming a better solution for IoMT security because it can learn features automatically and can understand complex patterns in large datasets. Models like DNN and Long Short-Term Memory networks often perform better than traditional ML models for detecting different cyber threats. Among these methods, DBNs are very important, since they are generative models built from stacked Restricted RBMs, and they use layer-by-layer training, which allows more stable performance, even in unsupervised situations [15].

By using many hidden layers, DBNs are able to take high-level abstract ideas from raw network data, and this can help more correctly classify malicious actions. However, the setup of the DBN structure, like how many layers and neurons should be used, is still not clearly agreed upon, and it remains an active research problem with no universal answer [6, 16].

The performance of these DL models strongly depends on the quality and also the variety of benchmarking datasets that are used for testing. The CICIoMT2024 dataset shows important progress because it gives a multi-protocol environment where traffic from MQTT, BLE, and Wi-Fi is included, and this reflects the mixed and complex nature of modern IoMT systems. Similarly, the ECU-IoHT dataset gives a more specific framework for studying cyberattacks such as Address Resolution Protocol (ARP) spoofing and Smurf attacks inside IoHT systems. Even though such datasets are available, many existing studies are still limited to only binary classification and do not give the needed multiclass detail, which is important for better threat reduction in healthcare networks [1, 17, 18].

Several studies demonstrated the effectiveness of DBNs for intrusion detection in IoMT and related IoT environments, consistently reporting high detection accuracy despite differences in network architectures, deployment settings and evaluation datasets [15, 19-21]. Early research established the feasibility of distributed DBN architectures for smart city IoT environments and multi-layer DBNs for wireless sensor networks, highlighting their scalability and strong detection performance [19, 20]. Subsequent work extended models to healthcare IoMT environments where DBNs achieved robust detection across diverse attack categories including Botnet, DoS/DDoS and Infiltration attacks, while further improvements were obtained through metaheuristic feature selection and hyperparameter optimization [1, 15]. Ensemble DBN architectures have also been investigated for cloud-supported IoT and SCADA systems, demonstrating that combining multiple DBNs can enhance adaptability and malware detection performance in industrial environments [21].

Beyond standalone DBN models, recent research has increasingly explored hybrid deep learning architectures to address the growing complexity of IoMT. These approaches integrate complementary learning paradigms, such as hierarchical meta learning for zeroday attack detection, recurrent neural networks for sequential traffic analysis and CNNs for extracting spatial temporal traffic features achieving detection accuracies approaching or exceeding 99% across diverse IoMT datasets [2, 8, 22]. Collectively, these findings suggest that while DBNs remain a competitive solution for intrusion detection, hybrid deep learning models can further improve feature representation and generalization, particularly in complex and heterogeneous IoMT environments.

Despite these advances, a critical pattern emerges: each study adopts a single fixed DBN or DL architecture and compares it against alternative model families such as SVM or CNN or RNN without systematically varying the internal architecture of the DBN. Key design variables like network depth, layer width, regularisation strategy and activation function are treated as constants rather than experimental factors, making it difficult to determine whether reported performance improvements stem from the choice of DL. Furthermore, several DBN studies rely on older non-IoMT-specific benchmarks such as NSL-KDD [22] and CICIDS2017 [15], which lack the multi-protocol heterogeneity and fine-grained attack taxonomy of modern IoMT datasets like CICIoMT2024 [17, 20] and ECU-IoHT [23]. Binary evaluation settings [8] further obscure class level weaknesses that are critical in healthcare security contexts where even a single missed attack category can have serious clinical consequences. These gaps motivate the present study's systematic and ablation style comparison of six DBN configurations across two complementary IoMT datasets.

1.3 Research gaps and study rationale

Despite recent advances, several important gaps remain in DBN intrusion detection for IoMT environments. First, most studies employ a single DBN architecture and compare it with other models such as SVM, CNN, or RNN while rarely investigating the DBN design space itself. Key architectural factors, including network depth, layer width, regularisation and activation functions are seldom examined systematically, making it difficult to identify the sources of performance improvements. Second, evaluations are often conducted on outdated or non-IoMT datasets such as NSL-KDD and CICIDS2017 or focus only on binary classification, failing to capture the multiclass, imbalanced and heterogeneous nature of real IoMT traffic. Third, cross-dataset validation is largely absent as models achieving high accuracy on a single dataset are rarely tested for generalisation across datasets with different characteristics. Finally, the impact of modern regularisation techniques, particularly Batch Normalisation and Dropout remains underexplored in DBN fine-tuning for IoMT intrusion detection [1, 2, 6, 15].

This study addresses these limitations by conducting a systematic evaluation of DBN architectures for IoMT intrusion detection. Six DBN configurations are designed to isolate the effects of depth (2 to 4 RBM layers), width (128 to 512 units), regularisation (Batch Normalisation and Dropout) and activation functions. The models are evaluated on two complementary datasets: Canadian Institute for Cybersecurity- Internet of Medical Things (CIC IoMT 2024) with 18-class high-dimensional benchmark, and ECU-IoHT with a compact 5-class dataset enabling cross-dataset generalisation analysis. The best-performing model is further validated through five independent runs with results reported as mean ± standard deviation to ensure reproducibility. By controlling architectural variables while keeping all other experimental conditions fixed, this work provides a rigorous ablation-style comparison of DBN design choices and addresses a key methodological gap in IoMT intrusion detection research.

3.1 Experimental overview

We conducted a comprehensive evaluation of DBN architectures for network intrusion detection across two distinct IoT healthcare traffic datasets. The two benchmarks differ in scale, dimensionality and classification complexity enabling cross-dataset assessment of DBN architectural choices. Table 3 shows a comparative overview between the datasets.

The identical experimental protocol is applied to both datasets. All DBN variants undergo 10 epochs of unsupervised RBM pretraining via Contrastive Divergence (CD-1) with a learning rate. $\eta=0.01$, followed by 100 epochs of supervised fine-tuning with the Adam optimiser under a cosine annealing schedule (initial $\eta=0.01$, $\eta_{\min}=10^{-4}$) and mini-batch size 2,048. Input features are standardised using StandardScaler. Six architectural variants are evaluated on both datasets.

A key distinction between the two benchmarks is the severity of the classification challenge: the CIC IoMT 2024 dataset presents 18 traffic classes across a 45-dimensional feature space with nearly 900 K test instances, while ECU-IoHT comprises 5 attack categories over a compact 10-feature space with 22 K test instances. As will be shown, this difference in intrinsic problem difficulty leads to qualitatively different behaviours across DBN architectures, providing complementary insight into the conditions under which each architectural choice is beneficial.

Table 3. Comparative overview of experimental datasets and evaluation conditions

3.2 Results on the Canadian Institute for Cybersecurity Internet of Medical Things 2024 dataset

3.2.1 Overview of classification performance

Table 4 presents the main evaluation metrics for all six DBN configurations on the CIC IoMT 2024 test set. For Experiment 1 and Experiment 2, an independent second training run was additionally conducted in order to quantify sensitivity to random initialization; these extra results are discussed within the respective experiment sections.

Table 4. Summary of classification performance across all six Deep Belief Network (DBN) experiments on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

The Regularised DBN (Experiment 5) is showing the highest overall performance, with a test accuracy is 97.18%, and macro-AUC is 0.9970. This result is considered stable in the evaluation stage. The ReLU DBN (Experiment 6) is almost the same in accuracy, about 97.15%, but AUC becomes lower at 0.9850; this means the probability output is less well calibrated, even if the final class decision is nearly identical, so some uncertainty quality is reduced.

The Standard DBN baseline (Experiment 2, main run) is recording the lowest test accuracy, 87.14%, and test loss is 0.5270, which is consistent with a training process that likely falls into not good local minimum in that particular run. However, another independent run of the same configuration is reaching 97.66% accuracy, and this suggests strong randomness in training outcome when regularisation is not applied, and performance is unstable across runs.

3.2.2 Experiment 1: Shallow Deep Belief Network

The Shallow DBN comprises two RBM pretraining layers (128 and 64 hidden units), the most compact architecture in the study. It achieves a test accuracy of 95.21% and a macro-AUC of 0.9887 in its primary run, as shown in Table 5.

Table 5. Shallow Deep Belief Network (DBN) per-class scores on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

A second independent training run produced substantially improved results: test accuracy of 97.87% and macro-AUC of 0.9952 (weighted F1 = 0.9760). The accuracy improvement of 2.66 percentage points between runs highlights the sensitivity of DBN training to random weight initialisation and the stochastic nature of CD-1 pretraining with a 45-feature input. The improved run shows markedly better precision and recall for Classes ‘DDoS ICMP’, ‘DDoS SYN’, ‘DDoS TCP’, and ‘benign’.

3.2.3 Experiment 2: Standard Deep Belief Network /baseline

The Standard DBN, which adds a third 128-unit hidden layer to the Shallow DBN, serves as the architectural reference baseline. In its primary run, it achieves 87.14% test accuracy and a macro-AUC of 0.9872, the lowest accuracy across all six experiments. As shown in Table 6.

Table 6. Experiment 2: Standard Deep Belief Network (DBN) per-class precision, recall, and F1-score on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

3.2.4 Experiment 3: Deep Deep Belief Network

The Deep DBN employs four RBM pretraining layers with a progressively narrowing architecture (256→128→64→32), culminating in a 32-unit final representation. It achieves 94.03% test accuracy and a macro-AUC of 0.9924, as can be seen in Table 7.

Relative to the Standard DBN primary run, the Deep DBN model shows an increase in test accuracy by 6.89% and also gives a slightly higher AUC by +0.0052; this is seen in Table 7.

Table 7. Experiment 3: Deep Deep Belief Network (DBN) per-class precision, recall, and F1-score on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

Class ‘DoS UDP’ has a great improvement, recall becomes 0.911 compared to 0.500 in Standard primary; however, Class ‘DoS ICMP’ recall is still very low at 0.087 and this is a critical problem. The gap between train and test accuracy is 2.63 percentage points (train: 96.66%, test: 94.03%), which can suggest moderate overfitting situation. The 32-unit bottleneck may cause compression risk in representation, and the still-low recall for Class ‘DoS ICMP’ and also Class ‘OS Scan’ (0.124) shows difficulty in keeping minority-class discriminative patterns through four sequential stochastic Bernoulli sampling stages.

3.2.5 Experiment 4: Wide Deep Belief Network

The Wide DBN quadruples the capacity of the first hidden layer (512 units versus 128 in the Standard DBN), maintaining a three-layer RBM stack. It achieves 94.69% test accuracy and macro-AUC of 0.9935.

Class ‘DoS Connect Flood’ achieves a perfect F1-score of 1.000, and Class ‘DoS UDP’ recovers substantially (F1 = 0.977, compared to 0.646 in the Standard DBN primary run) as shown in Table 8. However, the Wide DBN is unable to overcome the persistent weaknesses observed in Classes ‘DDoS Publish Flood’, ‘DoS ICMP’, ‘OS Scan’, and ‘Recon VulScan’.

Table 8. Experiment 4: Wide Deep Belief Network (DBN) scores on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

3.2.6 Experiment 5: Regularised Deep Belief Network

The Regularised DBN augments the Standard DBN topology with BN and Dropout applied after each hidden layer during supervised fine-tuning; the RBM pretraining phase is unchanged. It achieves the best overall performance across all primary runs: test accuracy of 97.18%, and macro-AUC of 0.9970, as shown in Table 9. To address run-to-run variability, this configuration was additionally evaluated across 5 independent training runs (seeds 1–5) using the same fixed train/test split. The repeated-run summary is: accuracy = 97.72 ± 0.04%, macro-AUC = 0.9940 ± 0.0023, and weighted F1 = 0.9742 ± 0.0004 (mean ± sample std). The per-class F1 mean ± std across the 5 runs is reported in Table 9.

Relative to the Standard DBN run, the Regularised DBN delivers marked per-class improvements: Class 1 F1 +0.142 (0.963 vs. 0.821), Class ‘DDoS TCP’ F1 +0.235 (0.919 vs. 0.684), Class ‘DoS ICMP’ F1 +0.664 (0.900 vs. 0.236, the most significant per-class improvement in the study), and Class ‘DoS UDP’ F1 +0.345 (0.991 vs. 0.646), confirming that BN and Dropout together steer the optimiser towards better-generalising solutions. All 18 classes exceed a per-class AUC of 0.980, indicating reliable discrimination in the probabilistic domain. Class ‘DDoS Publish Flood’ remains the most persistent challenge (recall = 0.138, F1 = 0.241), consistent across all models, suggesting that its feature distribution substantially overlaps with neighbouring classes in the 45-dimensional input space. Classes ‘OS Scan’ and ‘Recon VulScan’ also retain low F1-scores (0.163 and 0.209).

Table 9. Experiment 5: Regularized Deep Belief Network (DBN) performance on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024) (single primary run)

Figure 2 shows the per-class F1-score means and standard deviations across the 5 independent runs. The 5-run results confirm that the Regularised DBN performance is stable where the standard deviation of accuracy is only 0.045 percentage points across runs. Most classes exhibit tight F1 distributions (Table 10), though ‘Ping Sweep’ (std = 0.067) and ‘Recon VulScan’ (std = 0.032) show the highest relative variability, attributable to their small support sizes (169 and 973 instances, respectively).

Table 10. Per-class F1-score mean ± std for the Regularised Deep Belief Network (DBN) over 5 independent runs (seeds 1-5) on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

qq20260825-095803.png2_.png

Figure 2. F1 mean ± std bars with error caps for 5-run validation

3.2.7 Experiment 6: ReLU Deep Belief Network

The ReLU DBN replaces sigmoid activations in the supervised fine-tuning phase with Rectified Linear Units while retaining sigmoid activations during RBM pretraining achieving test accuracy of 97.15% and macro-AUC of 0.9850, as shown in Table 11.

Figure 3 shows the per-class precision, recall, and F1-score for the ReLU DBN on CIC IoMT 2024. The ReLU DBN matches the Regularized model closely in test accuracy (97.15% vs. 97.18%) and weighted F1-score (0.9686 vs. 0.9687), despite near-identical accuracy, reflecting the known tendency of ReLU networks trained with Adam to assign sharp, overconfident probability mass to dominant classes without correspondingly calibrating the full softmax distribution. This is evidenced by the AUC of 0.9850, the lowest in the results, and the lowest per-class AUC for Class ‘DDoS Publish Flood’ of 0.821, indicating that ReLU activations, while beneficial for decision-boundary sharpness, degrade probabilistic discrimination for inherently difficult classes.

3.png

Figure 3. Per-class precision/recall/F1 for ReLU Deep Belief Network (DBN)

Table 11. Experiment 6: ReLU Deep Belief Network (DBN) results on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

3.2.8 Comparative analysis

(1) Effect of network depth

On the contrary, the general expectation that deeper architectures learn more discriminative hierarchical representations, the Shallow DBN outperforms the Standard DBN by 8.07% in accuracy in the primary run comparison as can be seen in Table 12. The Deep DBN partially improves at 94.03%, exceeding the Standard DBN by 6.89%, but falls 1.18% points below the Shallow DBN. This non-monotonic depth-performance relationship is partly attributable to the training instability of the Standard DBN primary run; its second run achieves 97.66%, rendering the Deep DBN inferior by a further 3.63 percentage points. The AUC improvement of the Deep DBN over the Shallow DBN (+0.0037) is modest and does not translate to higher accuracy. These results indicate that for the CIC IoMT 2024 dataset, additional RBM depth beyond two layers does not yield consistent accuracy gains while incurring proportional training time increases.

Table 12. Classification performance as a function of Deep Belief Network (DBN) depth on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

(2) Effect of network width

The Wide DBN improves upon the Standard DBN by 7.55 percentage points in accuracy, as shown in Table 13, demonstrating that greater layer capacity assists convergence on the 45-feature, 18-class input. However, this improvement comes at a 2.14× training time cost, and relative to the Standard DBN accuracy (97.66%), the Wide DBN is inferior by 2.97 percentage points, indicating that wider layers do not overcome the fundamental convergence instability of unregularised DBN training. The Regularised DBN achieves superior accuracy (97.18%), confirming that regularisation is both more effective and more computationally efficient than brute-force width expansion.

Table 13. Classification performance as a function of first-layer width on CIC Internet of Medical Things (IoMT) 2024

(3) Effect of regularisation

As shown in Table 14, the Regularised DBN substantially outperforms the unregularised Standard DBN. Against the Standard DBN, the Regularised DBN delivers a 10.04 percentage point accuracy improvement, confirming that regularisation provides the convergence stability that the unregularised model achieves only unreliably.The macro-AUC of 0.9970 is the highest recorded in the study, attributable to Batch Normalisation's beneficial effect on class probability calibration.

Table 14. Effect of Batch Normalisation and dropout ($p=0.3$) on Deep Belief Network (DBN) performance on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

(4) Effect of activation function

The ReLU DBN achieves 97.15% accuracy compared to 87.14% in the Standard DBN (Table 15) a difference of 10.01 percentage points, reflecting confident class probability assignments characteristic of unregularised ReLU networks under Adam optimisation. ReLU is thus a cost-free accuracy improvement but sacrifices probabilistic discrimination quality relative to the regularised sigmoid model.

Table 15. Effect of activation function on Deep Belief Network (DBN) performance on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024)

3.2.9 Training dynamics and generalisation

Train-test accuracy gaps are uniformly modest across all six experiments (1.75–4.39 percentage points), indicating that no configuration is severely overfit in the classification sense. The Regularised DBN achieves the smallest accuracy gap (0.0183), confirming that BN and Dropout effectively constrain memorisation. Across 5 independent runs, the Regularised DBN’s train-test accuracy gap averages 0.0183 with negligible variation, demonstrating that BN and Dropout not only improve single-run generalisation but also eliminate run-to-run divergence. Table 16 summarises these convergence statistics for all six experiments.

Table 16. Training convergence statistics for all six experiments on Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024) (primary runs)

3.2.10 Class-level performance and persistent challenges

Three classes exhibit structural difficulty that no DBN variant resolves, as shown in Figure 4. Class ‘DDoS Publish Flood’ returns F1-scores of 0.218–0.252 with uniformly high precision (0.933–0.995) but extremely low recall (0.123–0.145), indicating near-complete misclassification of true instances into other categories; this pattern is consistent with substantial feature-space overlap between Class ‘DDoS Publish Flood’ and neighbours. Class ‘OS Scan’ achieves F1-scores of 0.142–0.208 with a recall below 0.124 in most experiments, despite a non-negligible support of 2,941. Class ‘Recon VulScan’ (973 test instances, the second-rarest class) reaches F1-scores of only 0.188–0.236, with both precision and recall in the low-to-mid range across all configurations.

In contrast, Class ‘DoS Publish Flood’ attains a stable F1 near 0.70 across all experiments by achieving a recall ≈ of 1.000 at the cost of a precision ≈ 0.54, a pattern invariant to architectural choice. Class ‘DoS Connect Flood’ performs near-perfectly (F1 ≥ 0.991) in all six experiments, indicating complete separability from all other classes. Detailed results for all the runs are presented in Table 17.

4.png

Figure 4. Per class F1-score for all experiments on CIC Internet of Medical Things (IoMT) 2024 as a bar chart representation

Table 17. Per-class F1-score across all six Canadian Institute for Cybersecurity- Internet of Medical Thing (CIC IoMT 2024) experiments (primary runs)

Table 28. Cross-dataset performance for the best and worst Deep Belief Network (DBN) configurations on each dataset

Note: Canadian Institute for Cybersecurity- Internet of Medical Things = CIC IoMT 2024; Edith Cowan University-Internet of Health Things = ECU-IoHT.

These findings suggest a dataset-dependent depth rule for DBN-based IoMT intrusion detection: use a shallow two-layer network when input features are few and classes are hard to separate; deeper networks help only when features are rich (e.g., 30+ features with 10+ classes). This aligns with prior work noting that optimal depth always depends on the dataset.

Width has opposite effects on different datasets. On ECU-IoHT (low-dimensional), a wide DBN achieves perfect accuracy, outperforming the standard version. On CIC IoMT 2024 (high-dimensional, multi-class), the wide DBN gets 94.69% accuracy. Expansion helps simple, low-dimensional problems but fails to scale to complex, high-dimensional ones, where regularization offers more benefit than simply adding capacity.

3.4.4 Training stochasticity and reproducibility

The two datasets exhibit fundamentally different training stability profiles. On ECU-IoHT run-to-run variability is low while the Standard DBN produces 252 misclassifications in its primary run; the architecture is consistent in its error pattern by always misclassifying No Attack as Smurf Attack. All six architectures that converge successfully do so to near-identical solutions with near-zero accuracy gaps. On CIC IoMT 2024, the Standard DBN primary and secondary runs differ by 8.52 percentage points in accuracy and the Shallow DBN two runs differ by 2.66 percentage points, a substantial variance for a deterministic architecture under fixed hyperparameters. This instability is a direct consequence of the larger and more complex problem landscape with 18 classes and 45 features; the CD-1 energy landscape contains more local minima, and the greedy layer-wise weight initialization is more sensitive to the random seed that determines the initial Boltzmann machine parameter values.

The practical implication is that on complex multi-class IoT traffic datasets, the unregularized DBN architectures require either multiple training runs with selection of the best outcome or the systematic application of Batch Normalization and Dropout to reduce sensitivity to initialization. Quantitatively, the Regularized DBNs’ accuracy standard deviation across 5 runs is only 0.045 percentage points on CIC IoMT 2024 and 0.000 on ECU-IoHT compared with an 8.52-percentage-point spread observed across two runs of the unregularised Standard DBN. The Regularized DBN thus eliminates this instability on both datasets making it the architecturally preferred choice for deployment.

3.4.5 ReLU activation-accuracy vs. calibration tradeoff

The ReLU DBN demonstrated a consistent cross-dataset pattern where it achieved near-optimal accuracy at negligible computational overhead over the sigmoid baseline, but sacrifices probabilistic calibration as measured by AUC. On ECU-IoHT, the ReLU model reduces errors by 99.2% and on CIC IoMT 2024, it nearly matches the Regularized DBN in accuracy within 0.03 percentage points, but records an AUC lower by 0.0120. For applications requiring calibrated probability outputs, such as risk stratification in healthcare IoT monitoring where continuous risk scores are clinically actionable, the regularized sigmoid architecture is preferred. For applications requiring only hard classification labels under tight computational budgets, the ReLU DBN provides an effective and efficient alternative.

3.5 Comparison with existing literature

The results are broadly consistent with prior research employing DBNs and other DL approaches for IoMT intrusion detection, while also revealing important differences in evaluation conditions and behaviors.

Most existing studies report very high accuracies in the range of 98%–99.9%. For instance, Jayanthi et al [1] achieved 98.71% using an optimized DBN with metaheuristic tuning, while Manimurugan et al. [15] reported up to 99.37% accuracy on the CICIDS 2017 dataset. Similarly, Otoum et al. [20] and Huda et al. [21]. reported near-perfect accuracies exceeding 99.7%. These results are comparable to the performance achieved in this work on the ECU-IoHT dataset where multiple DBN configurations reached 100% accuracy.

However, a key distinction lies in dataset complexity and evaluation realism. Many prior works rely on older or simplified benchmarks (e.g., NSL-KDD or CICIDS 2017) or binary classification settings, which are inherently less challenging. This study evaluates models on the more complex CIC IoMT 2024 dataset with18 classes, where the best achieved accuracy reached 97.18%. This gap highlights the increased difficulty of modern IoMT intrusion detection tasks and suggests that previously reported near-perfect results may not generalize to more realistic scenarios.

Recent non-DBN works support this observation. Uddin et al. [2] reported 99.77% accuracy on CICIoMT2024 using hierarchical meta-learning, while Mohammadi et al. [22] achieved 99% with a CNN model. Additionally, Algethami and Alshamrani [8] reported perfect performance on ECU-IoHT, consistent with the results observed in this work for low-dimensional datasets.

Overall, the comparison indicates that while DBNs remain competitive, particularly when properly regularized, their performance is sensitive to dataset complexity and architectural design.

3.6 Practical applicability and security engineering considerations

A security-focused view adds key issues for real-world DBN-based IDS in IoMT are false alarms, latency, drift resilience and feasibility.

False alarms matter in healthcare, where each alert requires clinician review and too many false ones cause fatigue. The regularised DBN hits high precision (≥0.90) in 14 of 18 CIC IoMT 2024 classes, but is weaker on 'DoS Publish Flood' (~0.54) and 'ARP Spoofing' (~0.40) creating extra triage load. On ECU-IoHT, its flawless precision removes this problem. Any real rollout needs either class-specific thresholds or verification for low-precision attacks.

The DBN's offline two-phase training is computationally expensive but acceptable for periodic updates. Inference is just a single forward pass through a few dense layers achieving sub-millisecond latency on modern edge hardware.

Concept drift and temporal robustness. IoMT traffic changes over time due to new devices, attacks and shifting patterns. Current DBN tests use static data missing this drift. A two-phase approach helps where unsupervised pretraining can update with new unlabeled data and supervised fine-tuning can adjust with small labeled batches.

Security engineering summary. The Regularised DBN suits IoMT with 97.72% accuracy, low false positives, fast inference and an adaptable architecture. However, three attack classes have high false positives, needing extra mitigation like human review.