Balancing the classes of an imbalance dataset is a crucial step to eliminate biased decisions caused by major classes over minority ones. In this work, SMOTE is applied only to the training subset in order to avoid information leakage and preserve the integrity of the validation and test distributions. As illustrated in Fig. 3, the process begins with loading the training set, then class distribution is examined to detect imbalance between majority and minority classes. The dataset is then separated into features x and labels y, and SMOTE is applied to generate synthetic samples for the minority classes. This step balances the dataset by ensuring that all classes are more evenly represented. The resampled features and labels are recombined into a new balanced dataset, and the class distribution is rechecked to confirm the effect of SMOTE. Finally, the balanced dataset is saved as a CSV file for use in subsequent model training. This process ensures that the learning algorithm does not become biased toward the majority class and can achieve more reliable and fair performance across all classes.

MI between a feature Xj and the label Y is: (1)I(Xj;Y)=∑xj∈Xj∑y∈Yp(xj,y)log⁡p(xj,y)p(xj)p(y)

Eq. (1) defines the standard MI feature selection, which measures how much knowing a feature Xj helps in predicting the class label Y. In simple terms, MI evaluates the dependency between a feature and the label. The term p(xj,y) represents the joint probability of the feature value and the label occurring together. However, p(xj) and p(y) represent their individual probabilities. Therefore, Xj and Y are independent, then p(xj,y)=p(xj)p(y), and the logarithmic term becomes zero. This means that the feature provides no useful information about the label. Larger MI values indicate that the feature carries more information about the label, making it more valuable for classification tasks. Features are ranked by I(Xj;Y) and the top-K are selected.

The cumulative fraction policy (cumfrac) in Eq. (2) selects the smallest number of top features K such that their combined MI accounts for at least 90% of the total MI across all features. In practice, this means we first sort features by how much information they share with the label. Then we keep adding features, one by one, until the fraction of total information they explain reaches the threshold τ=0.90. This ensures that most of the useful information is kept while removing redundant or less important features [38]. (2)K=min{k:∑i=1kI(X(i);Y)∑j=1nI(X(j);Y)≥τ},τ=0.90where X(i) are features sorted by decreasing MI score.

The framework employs two random forest classifiers, which are a base model and a reduced model. The base model uses the entire feature set with dimensionality d, whereas the reduced model is restricted to only the top-K most informative features selected using MI. This allows the reduced model to achieve faster predictions with lower computational cost, while still maintaining competitive accuracy. The mathematical representations of the two models are given below:

The base model uses all d features is given in Eq. (3): (3)y^base=fRF(X),X∈Rd

While the reduced model uses the top-K features is represented in Eq. (4): (4)y^reduced=fRF(X1:K),K≪d

Fig. 4 demonstrates the workflow of the base and reduced models. The base model directly uses all available input features, training a random forest followed by probability calibration without applying feature selection.

In contrast, the reduced model applies MI feature ranking to the input features, retaining the top-K subset based on a cumulative fraction threshold of τ=0.90, which ensures that at least 90% of the overall information content is preserved. A random forest classifier is then trained on this reduced feature set. This dual setup, one model using all features and another using only the most informative subset, allows a clear comparison of their performance. This shows how the MI method helps remove redundancy, reduce dimensionality, and minimize noise.

Although IoT features are numeric, transforming each feature vector into a tokenized text sequence allows ELECTRA to model dependencies between features that traditional tabular methods treat independently. The transformer captures contextual relationships, feature combinations, and non-linear interactions that often characterize malicious IoT traffic. Using a lightweight logistic regression probe on top of these embeddings provides an efficient refinement step when the base model is uncertain, enabling richer representation learning with minimal additional cost. As shown in Eq. (5), each feature vector Xi is first converted into a text string and then passed through the ELECTRA model, which produces a dense embedding zi∈Rh. These embeddings capture richer semantic and structural patterns from the input features, which effectively maps raw numerical data into a representation space where complex relationships are easier to learn.

Features are transformed to text and embedded with ELECTRA: (5)zi=ELECTRA(Tokenizer(features(Xi))),zi∈Rh

After the embeddings are created the Teacher Probe is implemented as a logistic regression classifier as represented in (6). The probe is trained on embeddings derived from the training set, making it computationally efficient while still leveraging the representational power of ELECTRA. At test time, when the base random forest model is uncertain, the Teacher Probe predicts labels using the test embeddings. This combination allows the framework to integrate deep language model representations into a classical ML pipeline without requiring fine-tuning of the large transformer itself.

A teacher probe (logistic regression) is trained on embeddings: (6)y^probe=fLR(zi)

Entropy is used to to measure the uncertainty of a model or dataset [25]. Given a probability distribution p=(p1,p2,…,pi) over i classes, the entropy as shown in Eq. (7), H(p) calculates how confused the model is. If the distribution is sharp for example, one class has probability close to 1, the entropy is low, indicating high confidence. While, if the probabilities are not sharp, the entropy is higher, meaning the model is uncertain. (7)H(p)=−∑i=1npilog⁡(pi),where pi is the predicted probability for class i.

To decide when to trust the base model and when to move to a more expensive option (such as the Teacher Probe), an adaptive threshold T is used as in (8). This threshold is computed using the z-score of the entropy values across a batch. Specifically, the dynamic threshold is set to the mean entropy µH plus a multiple λ of the standard deviation σH. (8)T=µH(p)+λ⋅σH(p),where µH(p) and σH(p) are the mean and standard deviation of entropy values, and λ is a scaling factor.

The agent has four possible actions, each representing a different trade-off between accuracy and computational cost.

Reduced Model: The first action is using random forest that is trained only on the top-K features selected using the MI cumfrac policy. By focusing on the most informative features, the reduced model is faster and simpler, with no added cost. However, it may sacrifice some accuracy compared to the full model. Reduced Features action can be represented as in Eq. (9) (9)y^=fred(X1:K)

Base Model: This is the second action as defined in (10), which uses the calibrated random forest model trained on all d features. It provides strong predictive performance since no information is discarded. Importantly, this option is considered cost-free, making it a reliable baseline. (10)y^=fbase(X)

LLM Refinement: In this third action, the Base Model initially performs the prediction, and its confidence is quantified using the entropy of the predicted probability distribution, denoted as H(p). If the entropy is below the adaptive threshold T (indicating low uncertainty), the prediction from the Base Model is accepted directly. Conversely, if the entropy exceeds the threshold (indicating high uncertainty), the system activates the Teacher Probe, a logistic regression classifier trained on ELECTRA embeddings. This mechanism refines the decision for ambiguous samples by leveraging deep contextual representations. The decision rule for this action is formally expressed in Eq. (11): (11)y^i={fbase(Xi),if H(pi)<T,fprobe(zi),if H(pi)≥T,

Escalate to Analyst: The fourth action is used when predictions are highly uncertain. The top K samples with the highest entropy values are forwarded to an analyst. In practice, this means that the system does not assign a predicted label for these cases and instead marks them as unresolved pending manual verification. As shown in Eq. (12), if the sample i is not in the Top-K uncertain samples (i∉TopK(H(p))), the prediction is made by the base model fbase(Xi). However, if sample i is in the Top-K uncertain samples (i∈TopK(H(p))), the output is assigned an UNRESOLVED label to reflect escalation to the analyst. (12)y^i={fbase(Xi),if i∉TopK(H(p)),UNRESOLVED,if i∈TopK(H(p)).

Fig. 5 illustrates the workflow of the adaptive decision approach, which controls the dynamic action selection process during inference. The input batch, which comes from the test set, is passed to MI model to perform feature selection using a cumulative fraction policy to identify the most informative subset of features. This step ensures that redundant or less relevant features are excluded.

The selected features are then submitted to the Base Model to produce class predictions along with confidence scores. These scores are checked using an adaptive entropy method to decide how reliable each prediction is.

As the figure shows, the model dynamically selects one of four decision pathways based on the evaluated entropy level. For instance, in case of low-entropy, the Reduced Model path is used after selecting the top-K most informative features. It is worth noting that this option is computationally efficient and suitable when predictions are already confident. The system switches to the Base Model when the entropy level increased to operate on the complete feature set. This enhances reliability in uncertain cases. The LLM Refinement process is activated by the framework when the entropy rises. This uses deep contextual representations together with logistic regression and ELECTRA embeddings to enhance predictions.

Finally, the agent initiates the Escalate to Analyst pathway when the entropy stays over a certain threshold. For manual verification, it sends the top-K highly uncertain instances. The process concludes at the Accept Prediction module where the result of the action selected is used to make the final decision of the system. This adaptive control improves accuracy, interpretability and computational efficiency by intelligent allocation of computational resources based on the uncertainty of each input sample.

Overall, these four steps allow the model to effectively balance computational costs toward predictive performance. In order to maintain efficiency without reducing accuracy, the framework is encouraged by the cost structure to keep the more costly choices (Actions 3 and 4) to only highly uncertain samples.

The reward function is designed to balance prediction accuracy with the computational or operational penalty of each action. The reward calculates the difference between the performance score of the model and the penalty of the chosen action as defined in Eq. (13). We use accuracy as the scoring function that measures the proportion of correctly classified samples. Actions that achieve high accuracy with minimal penalty will receive higher rewards. However, actions that are more expensive or less accurate will produce lower rewards. For instance, there is no penalty associated with the usage of the Reduced or Base models, therefore the accuracy is equivalent to the reward. By contrast, utilizing the Teacher Probe or forwarding the Top-K uncertain samples to the analyst incurs additional penalties, which reduces the reward even when accuracy is high. This design ensures that the system learns to prefer strategies that are not only correct but also efficient, reflecting a realistic trade-off between accuracy and resource usage. (13)r=Score(y,y^)−Penalty(a)

Each action in the LinUCB algorithm keeps its own parameters, which are represented as a vector ba and a matrix Aa. The matrix Aa∈Rd×d captures how much information has been gathered about the action based on past feature contexts, while ba∈Rd accumulates the rewards observed [39]. Using these values, the algorithm estimates the parameter vector θa=Aa−1ba, which can be interpreted as the model’s best guess about the relationship between the context and the expected reward for action a. When deciding which action to take, LinUCB computes an upper confidence bound (UCB) score for each action as written in Eq. (14): (14)pa(x)=θa⊤x+αx⊤Aa−1x.

The first term, θa⊤x, represents the predicted reward for the action given the current context (features) x. The second term adds an exploration bonus, scaled by the parameter α. This bonus is larger when the algorithm has less certainty about the action (i.e., when Aa−1 is large), encouraging exploration of under-tested actions.

The action with the highest score is then selected as shown in Eq. (15): (15)at=arg⁡maxapa(xt),where pa(xt) represents the estimated upper confidence bound for action a given context xt. This means the agent selects the action that maximizes the combination of the predicted reward and the uncertainty bonus.

Once the reward rt is observed, the algorithm updates its internal parameters for the chosen action as in Eq. (16): (16)Aa←Aa+xtxt⊤,ba←ba+rtxt.

These updates allow LinUCB to continuously refine its confidence estimates, gradually learning which actions yield higher rewards under different contexts.

In this research, we evaluated our proposed adaptive intrusion detection framework on three benchmark datasets including RT-IoT-2022 [40], CIC-IoT-2023 [41] and CIC-IoMT-2024 [42]. We chose these datasets because they represent modern network behavior, diverse device types and unseen attacks in IoT and IoMT environments.

The first dataset used is RT-IoT-2022 and contains attack and normal traffic from a real-time IoT setup. It was designed for intrusion and anomaly detection research and provides 83 flow-based features such as packet statistics, payload details, timing and TCP flags [40]. The dataset has nine attack types that are SYN flooding, Slowloris distributed denial of service (DDoS), ARP poisoning and Nmap scans as well as three types of benign traffic. This variety makes this dataset a strong test for classification methods, particularly given the limited resources common in IoT systems.

The second dataset is CIC-IoT-2023 that also supports research on intrusion and anomaly detection and was developed using a network of 105 real IoT devices [41]. It has 33 attack types from compromised devices that are grouped into seven categories namely DDoS, DoS, reconnaissance, web attacks, brute force, spoofing and Mirai botnet variants. The dataset provides 46 features describing network flows, packet statistics and timing. This range enables us to evaluate our detection methods on a variety of new IoT attacks and takes various devices and protocols into account.

The third dataset is CIC-IoMT-2024 that was collected from a simulated healthcare network with typical medical IoT devices and protocols [42]. The dataset includes both benign and malicious traffic such as resource loss attacks, illegal access, and probing. It provides 44 flow-based features on protocol use, packet statistics and timing. This setup lets us test our detection methods in healthcare where device variety and the need for constant service add extra challenges.

Overall, the use of the aforementioned datasets to evaluate our intrusion detection framework helps to demonstrate its effectiveness and adaptability to diverse IoT and IoMT environments. The variety of attack types, device behaviors and protocol characteristics in RT-IoT-2022, CIC-IoT-2023, and CIC-IoMT-2024 enables a thorough evaluation of the classification performance of our model under realistic conditions. Table 2 summarizes the key characteristics of the datasets used in this study, highlighting their differences in environment, feature dimensionality, attack diversity, and application domain.

MI was employed to estimate the relevance of each feature to the target class and to guide dimensionality reduction. Fig. 6 shows the ranked MI scores for all features in the RT-IoT-2022 dataset. Each bar represents a feature’s MI value, where higher scores indicate stronger dependence on the class label. Using the cumulative-fraction policy with a threshold of τ=0.90, the cutoff indicated by the red dashed line retains K = 58 out of d = 82 features. These 58 features account for at least 90% of the total MI, while the remaining variables contribute only marginal information and are removed. This reduction eliminates redundant or low-impact features and lowers computational cost by focusing the model on the most informative subset. To justify τ=0.90, we evaluated thresholds τ = 0.70, 0.80, 0.90, 0.95, and 0.99 as presented in Table 3. The MI values drop quickly, so a small number of features carry most of the useful information. At τ = 0.70 and τ = 0.80, only 41 and 49 features are needed to cover most of the MI. The reference threshold τ = 0.90 selects 58 features, providing a stable balance between information coverage and dimensionality. Increasing the threshold beyond this point brings only minor benefit. For example, τ = 0.99 increases the count to 72 features, but the extra features add almost no meaningful information.

Jaccard similarity analysis also supports this choice. When using τ=0.70 or τ=0.80, the selected features are simply smaller versions of the same important set obtained at τ=0.90. When increasing the threshold to τ=0.95 or τ=0.99, the extra features added are mostly low-importance variables that do not change the core set. In short, τ=0.90 preserves the meaningful features while avoiding unnecessary, low-impact variables.

The percentage of each action is calculated to determine action usage in the RT-IoT-2022 dataset. This helps identify which action dominates when the adaptive framework is used. As shown in Fig. 7a, the LinUCB controller can dynamically select among four possible inference actions based on prediction confidence and computational cost. There are four actions as mentioned in the methodology section named Reduced Model, Base Model, LLM Refinement, and Escalate to Analyst. The overall results demonstrate that the Reduced Model and Base Model together recorded 50% of all decisions. The LLM Refinement was used in 27.55% of the total actions. The Escalate to Analyst accounted for 22.45%, which is desirable since relying less on this costly action improves overall efficiency.

Therefore, the proposed framework reduces computation, as half of the actions involve no extra processing cost. The results also show that the Reduced Model being the most frequently chosen 27.93%, which reflects the system’s preference for computational efficiency inference when confidence is high.

The per-action performance on the RT-IoT-2022 dataset using 5-fold average is shown in Table 4. The Reduced Model and Base Model achieve the best overall balance, with high accuracy at 0.9992 and 0.9986, zero penalty, and the lowest inference time per sample at 0.105 and 0.107 ms. These results confirm that lightweight actions handle most samples efficiently without compromising accuracy. For uncertain cases, the framework triggers LLM Refinement. ELECTRA reaches 0.9962 accuracy and 0.9981 F1-score, while DistilBERT achieves 0.9973 accuracy and 0.9986 F1-score. The two models deliver comparable refinement performance, with only minor differences in accuracy and F1, and both incur the same five percent penalty and inference time of about 0.238 ms per sample. Escalate to Analyst achieves the highest accuracy at 0.9999 and F1-score at 0.9948 but carries the highest penalty (10%) and therefore the lowest reward among the correct actions. Overall, the results show that the framework prioritizes fast, zero-penalty actions and uses costly steps only when uncertainty is high.

It is worth mentioning that each experiment uses 5-fold cross-validation. We used 5-fold cross-validation to balance statistical robustness and computational efficiency, as higher fold counts (e.g., 10-fold) substantially increase training and evaluation cost and are commonly reported to provide only marginal performance gains when sufficient data is available.

In each fold, the dataset is first split into 80% training and 20% testing. The training portion is then further divided into 80% training and 20% validation, resulting in final proportions of 64% training, 16% validation, and 20% testing. The validation set is used for model calibration, feature selection, and adaptive decision tuning, while the test set remains completely unseen until final evaluation.

Different split ratios mainly affect the bias–variance trade-off. Increasing the training portion may slightly improve fitting but reduces the reliability of validation and test estimates, whereas smaller training sets increase variance, particularly in imbalanced IoT traffic. The chosen configuration yields stable and consistent performance across folds.

We also analyzed the trade-off between accuracy and the action penalty across the four decision actions of the proposed framework on the RT-IoT-2022 dataset. The x-axis represents the assigned penalty, while the y-axis shows the average accuracy, as illustrated in Fig. 7b. The optimal region appears on the far left, where high accuracy aligns with minimal penalty. Conversely, the far-right region reflects actions that achieve high accuracy but incur higher penalties.The Reduced Model and Base Model appear in the far-left region and achieve near-perfect accuracy with zero penalty, confirming their efficiency for standard predictions. The LLM Refinement action introduces a moderate penalty with a slight decrease in accuracy, and is triggered only for uncertain samples. The Escalate to Analyst action, which offers the highest accuracy, carries the highest penalty and appears in the upper-right corner, reflecting the cost of manual verification. Overall, the figure shows that the framework balances accuracy and efficiency. Zero-penalty actions handle most inputs reliably, while higher-penalty actions are used only when uncertainty is high.

The MI distribution for the CIC-IoT-2023 dataset, which contains d=46 features. At this point, K=23 features are retained, while the remaining 23 features, which contribute only marginally to the predictive signal, are excluded. This selection balances dimensionality reduction with information retention, ensuring that the model leverages the most informative subset of features while discarding less relevant ones to reduce complexity and noise.

The MI distribution of the CICIoT-2023 dataset reveals that among the total d=46 extracted features, the informative content is highly concentrated within the top-ranked subset. As presented in Fig. 8, the features are arranged in descending order of their MI scores, showing a rapid decline after the leading features. This indicates that only a limited number of them significantly contribute to class discrimination. Using the cumulative fraction policy with a threshold, a total of K=23 features are retained, capturing 90% of the total MI, while the remaining features are discarded due to their marginal relevance.

The distribution of action usage within the proposed framework on the CIC-IoT-2023 dataset is analyzed in this research. Fig. 9a illustrates that the Reduced Model and Base Model account for around 37.52% and 36.70% of all forecasts, respectively and therefore lead the decision-making process. This highlights the preference of the framework for low-cost inference pathways, which confirms that most network traffic can be accurately classified using lightweight or full-feature models without the need for additional computation. The LLM Refinement action, which involves deeper semantic processing through ELECTRA embeddings and logistic regression, is used for only 9.72% of samples. This indicates that uncertainty levels in the baseline predictions are relatively low and the refinement mechanism is triggered selectively for uncertain cases. Finally, the Escalate to Analyst action was used for 16.06% of the cases, which represents that a small portion of samples had the highest prediction entropy. Overall, the adaptive framework successfully used the Reduced and Base Models for 74.22% of the cases and only about 25.78% were handled by LLM Refinement and Escalate to Analyst. This emphasizes that the model primarily relies on low-cost actions.

On the CIC-IoT-2023 dataset,the Reduced Model and Base Model deliver almost identical performance with accuracies of 0.9417 and 0.9416 and very low inference time in the range of 0.1020 to 0.1110 ms as reported in Table 5. This indicates that the framework maintains strong predictive capability under both lightweight and full-feature inference without any penalty. The LLM Refinement actions powered by ELECTRA and DistilBERT produce closely matched results with accuracies near 0.916 to 0.919 and weighted F1-scores near 0.951 to 0.953. Both actions incur a five percent penalty, leading to rewards in the range of 0.865 to 0.869, and introduce a slightly higher latency of about 0.24 ms. The Escalate to Analyst action achieves the highest accuracy of 0.9459 and a weighted F1-score of 0.9698, though its ten percent penalty reduces the reward to 0.8459. Overall, the controller maintains a stable detection performance while using higher-cost actions only when uncertainty makes them necessary.

Fig. 9b shows the trade-off between accuracy and action penalty on the CIC-IoT-2023 dataset. The Reduced Model and Base Model appear in the top-left region, reaching about 0.94 accuracy with zero penalty, confirming their efficiency for most samples. LLM Refinement lies in the middle with a 5% penalty and lower accuracy (around 0.92), reflecting its selective use for uncertain cases. Escalate to Analyst action achieves accuracy around 0.945 but at the highest penalty (10%), indicating that escalation is reserved for the most difficult, high-risk samples.

Fig. 10 presents the MI distribution for the CIC-IoMT-2024 dataset, which has d=45 features. The blue bars correspond to MI scores and the solid line represents cumulative MI. The use of the τ=0.90 cutoff, K=25 features are selected, as shown by the vertical dashed red line. The retained subset accounts for at least 90% of the overall information, while 20 less-informative features are removed. This selection strategy reduces computational complexity while preserving the predictive richness of the dataset.

The distribution of action usage for the proposed framework in the CIC-IoMT-2024 dataset is shown in Fig. 11a. The results show that the Reduced Model and Base Model together dominate the adaptive decisions as they accounted for 68.75% of all predictions. This reflects the framework’s strong preference for computationally efficient because most samples are confidently classified without the need for additional processing. The LLM Refinement action was triggered for 15.48% of instances, which represents uncertain samples where the confidence of the primary models falls below the adaptive entropy threshold. Meanwhile, the Escalate to Analyst action was invoked for 15.77% of the inputs, which corresponds to the most complex samples with the highest uncertainty levels. These cases are routed for expert-level or external analysis to ensure accuracy and reliability. The near-balanced distribution between low-cost and high-cost actions confirms that the adaptive policy dynamically optimizes the trade-off between efficiency and precision, which ensures that the system maintains high detection accuracy while minimizing unnecessary computation.

The framework was also evaluated on CIC-IoMT-2024 dataset using 5-fold cross-validation, and the averaged per-action results as shown in Table 6. The Reduced Model and Base Model actions provide the strongest overall balance, reaching accuracies of 0.9971 and 0.9976 with weighted F1-scores of 0.9984 and 0.9987, zero penalty, and small latency of about 0.105 to 0.109 ms.

The LLM Refinement action achieves accuracies of 0.9842 with ELECTRA and 0.9835 with DistilBERT, with corresponding weighted F1-scores of 0.9909 and 0.9909. Both models deliver comparable accuracy, and the five percent penalty reduces the rewards to 0.9342 and 0.9335. Latency increases slightly due to the refinement step, reaching roughly 0.239 to 0.240 ms per sample. The Escalate to Analyst action attains 0.9915 accuracy and 0.8999 F1-score but carries the highest ten percent penalty.

The relationship between accuracy and action penalty on the CIC IoMT 2024 dataset is shown in Fig. 11b. The Reduced Model and Base Model actions appear at the zero penalty region with accuracies of 0.9971 and 0.9976. They deliver the strongest efficiency to accuracy value and remain the dominant choices for most traffic. The LLM Refinement action carries a five percent penalty and achieves an accuracy of about 0.984. This reduction is expected because the action is triggered only for high-uncertainty samples, which are naturally more difficult to classify. The Escalate to Analyst action appears at the ten percent penalty level and reaches an accuracy of about 0.991. While not the highest, it provides the most reliable option for the hardest cases because it represents human verification.

To further validate the findings of our research, we compare the classification accuracy and F1-score of the proposed framework with existing intrusion detection approaches, as summarized in Table 7. Unlike prior studies that report a single static model, the proposed framework is adaptive and operates through multiple inference paths. Depending on prediction uncertainty and computational constraints, the controller dynamically selects between lightweight (Reduced Model), full feature (Base Model), refinement, or escalation strategies. Therefore, Table 7 reports the performance of the proposed framework under its two primary zero-penalty inference paths (Reduced and Base Models), which represent the dominant operating modes during deployment.

Hafid et al. [35] reported strong performance on CIC-IoMT-2024 using a late-fusion XGBoost and Logistic Regression model, achieving 97% accuracy and 98% F1-score. While both their approach and the proposed framework incorporate adaptivity, their interpretability relies on SHAP values and their evaluation follows a static cost-performance analysis. Their method reduces bandwidth usage by 73% with detection latency below 200 ms, whereas our framework achieves substantially lower inference time in the sub-millisecond range.

Houichi et al. [32] reported 99% accuracy and a lower F1-score of 91% on CIC-IoT-2023 using traditional ML classifiers. However, their evaluation was limited to a single dataset and lacked adaptive decision control, limiting generalizability.

Prasad et al. [27] achieved 98% accuracy and a 98% F1-score on CIC-IoT-2023 using a DL model with cosine-similarity–based balancing. Their balancing method mitigated the inherent class imbalance of the dataset. In addition, their deep-learning pipeline followed a static inference process with no adaptive or cost-aware control. In contrast, the proposed adaptive framework maintains consistently strong performance across three benchmark datasets (RT-IoT-2022, CIC-IoT-2023, CIC-IoMT-2024) and introduces a LinUCB-based decision layer that dynamically balances accuracy and penalty under varying network conditions. They reported a 53% reduction in training time, whereas our framework delivers fast inference with an average cost near 0.1 ms per sample.

Almohaimeed and Albalwy [24] applied multiple feature selection methods to RT-IoT-2022 followed by an MLP classifier, achieving 96% accuracy and a 92% F1-score. While effective at reducing dimensionality, their approach does not incorporate adaptive or cost-aware decision logic. Our framework surpasses these results, achieving up to 99.92% accuracy and 99.95% F1-score on RT-IoT-2022 using the Reduced Model. Their method reduced runtime by 66.4%, whereas our framework additionally achieves sub-millisecond inference with higher predictive performance.

Finally, Mahdi et al. [31] used the CIC-IoT-2023 dataset, which is inherently imbalanced, with attack traffic dominating normal flows. The authors did not apply any explicit data-balancing method, instead relying on the incremental learning process to adapt to changing distributions. Consequently, the reported 99.53% accuracy may be influenced by class imbalance, as no F1-score or weighted metrics were provided. Their system introduces blockchain-related overhead (0.0001 s–0.0003 s per operation and 2–3 s per transaction), whereas the proposed IDS maintains inference latency near 0.1 ms.