• Developed RAID-KL, a teacher-student framework tailored for IoT security. • Validated RAID-KL on multiple data, achieving superior accuracy and efficiency. • RAID-KL achieves 11.3% and 64.33% reduction in CPU and memory usage, respectively. • RAID-KL compressed teacher model by 91.24% with an accuracy of approx. 99.75% • Applied SHAP to interpret feature contributions for IoT attack detection. Knowledge distillation (KD) is a pivotal model compression technique that enables the deployment of lightweight neural networks on resource-constrained Internet of Things (IoT) devices without sacrificing predictive performance. In the literature, KD has been widely applied to intrusion detection designs using the Kullback-Leibler (KL) divergence loss to align the distillation loss. However, KL divergence suffers from asymmetry and numerical instability when student predictions poorly match teacher distributions. The Jensen-Shannon (JS) divergence presents an alternative that allows both distributions to be treated equally relative to their average and ensures more balanced knowledge transfer process. The JS although symmetric and bounded, may converge more slowly due to its conservative nature. This study introduces RAID-KL , a resource-aware security model that leverages KD and a novel adaptive loss function combining hybrid KL and JS divergence. By integrating the benefits of both divergences, our method enhances the generalization and convergence of distilled models while maintaining low computational overhead, balanced knowledge sharing and improved numerical robustness. RAID-KL utilizes 1D Convolutional Neural Networks (1DCNNs) as the learning algorithm with teacher-student paradigm where a complex model serves as the teacher model, transferring its learned representations to a significantly smaller student network. RAID-KL is trained and evaluated in real-world network traffic datasets, including CICIoT2023, CICIoMT2024 and NIMSLABIoT2025, which include several IoT threats. RAID-KL demonstrates high performance in all applied metrics and low resource utilization during training and inference. To elucidate the critical relationships and feature contributions in model decisions, we integrate SHapley Additive exPlanations (SHAP) values for interpretability. Our findings reveal that the choice of loss function significantly impacts both performance and resource efficiency, with the hybrid KL-JS loss achieving superior trade-offs. Empirically, RAID-KL achieves 11.3% reduction in CPU usage and 64.33% reduction in memory usage during inference. Additionally, the RAID-KL model achieves 91.24% model compression over the teacher model while maintaining nearly the same classification accuracy. These insights highlight the robustness of RAID-KL framework, while providing a nuanced explainable index compared to the literature.

The integration of artificial intelligence (AI) into medical image analysis has transformed healthcare, offering unprecedented precision in diagnosis, treatment planning, and disease monitoring. However, its adoption within the Internet of Medical Things (IoMT) raises significant challenges related to transparency, trustworthiness, and security. This paper introduces a novel Explainable AI (XAI) framework tailored for Medical Cyber-Physical Systems (MCPS), addressing these challenges by combining deep neural networks with symbolic knowledge reasoning to deliver clinically interpretable insights. The framework incorporates an Enhanced Dynamic Confidence-Weighted Attention (Enhanced DCWA) mechanism, which improves interpretability and robustness by dynamically refining attention maps through adaptive normalization and multi-level confidence weighting. Additionally, a Resilient Observability and Detection Engine (RODE) leverages sparse observability principles to detect and mitigate adversarial threats, ensuring reliable performance in dynamic IoMT environments. Evaluations conducted on benchmark datasets, including CheXpert, RSNA Pneumonia Detection Challenge, and NIH Chest X-ray Dataset, demonstrate significant advancements in classification accuracy, adversarial robustness, and explainability. The framework achieves a 15% increase in lesion classification accuracy, a 30% reduction in robustness loss, and a 20% improvement in the Explainability Index compared to state-of-the-art methods.

Abstract With the substantial industrial growth, the industrial internet of things (IIoT) and many IoT avenues have emerged. However, the existing industrial architectures are still inefficient to deal with advanced security issues due to the distributed and distensible nature of the network IIoT communication networks. Therefore, solutions for improving intelligent decision‐making actions to the IIoT are sorely necessary. Thus, in this paper, the main cybersecurity attacks are predicted by applying a deep learning model. The various security and integrity features such as the DoS, malevolent operation, data type probing, spying, scanning, intrusion detection, brute force, web attacks, and wrong setup is analysed and detected by a novel sparse evolutionary training (SET) based prediction model. To scrutinize the conduct of the proposed SET‐based prediction model, evaluation parameters, such as, precision, accuracy, recall, and F1 score are measured and compared to other state‐of‐the‐art algorithms, in which the proposed SET‐based model achieved an average accuracy of 0.99% for an average testing time of 2.29 ms. Results reveal that the proposed model improved the attack detection accuracy by an average of 6.25% when compared with the other state‐of‐the‐art machine learning models in a real scenario of IoT security in Industry 4.0.

Light-field (LF) cameras, also known as plenoptic cameras, permit the recording of the 4D LF distribution of target scenes. However, many times, surface errors of a microlens array (MLA) are responsible for degradation in the images captured by a plenoptic camera. Additionally, the limited pixel count of the sensor can cause missing parallax information. The aforementioned issues are crucial for creating accurate maps for Intelligent Autonomous Transport System (IATS), because they cause loss of LF information, and need to be addressed. To tackle this problem, a learning-based framework by directly simulating the LF distribution is proposed. A high-dimensional convolution layer with densely sampled LFs in 4D space and considering a soft activation function based on ReLU segmentation correction is used to generate a superresolution (SR) LF image, improving the convergence rate in the deep learning network. Experimental results show that our proposed LF image reconstruction framework outperforms the existing state-of-the-art approaches; specifically, it is effective for learning the LF distribution and generating high-quality LF images. Different image quality assessment methods are used to evaluate the performance of the proposed framework, such as PSNR, SSIM, IWSSIM, FSIM, GFM, MDFM, and HDR-VDP. Additionally, the computational efficiency was evaluated in terms of number of parameters and FLOPs, and experimental results demonstrated that our proposed framework reached the highest performance in most of the datasets used.

Telecommunication service providers have to guarantee acceptable speech quality during a phone call to avoid a negative impact on the users’ quality of experience. Currently, there are different speech quality assessment methods. ITU-T Recommendation G.107 describes the E-model algorithm, which is a computational model developed for network planning purposes focused on narrowband (NB) networks. Later, ITU-T Recommendations G.107.1 and G.107.2 were developed for wideband (WB) and fullband (FB) networks. These algorithms use different impairment factors, each one related to different speech communication steps. However, the NB, WB, and FB E-model algorithms do not consider wireless techniques used in these networks, such as Multiple-Input-Multiple-Output (MIMO) systems, which are used to improve the communication system robustness in the presence of different types of wireless channel degradation. In this context, the main objective of this study is to propose a general methodology to incorporate wireless network parameters into the NB and WB E-model algorithms. To accomplish this goal, MIMO and wireless channel parameters are incorporated into the E-model algorithms, specifically into the $I_{e,eff}$ and $I_{e,eff,WB}$ impairment factors. For performance validation, subjective tests were carried out, and the proposed methodology reached a Pearson correlation coefficient (PCC) and a root mean square error (RMSE) of 0.9732 and 0.2351, respectively. It is noteworthy that our proposed methodology does not affect the rest of the E-model input parameters, and it intends to be useful for wireless network planning in speech communication services.

In communication services, speech quality plays an important role to achieve user expectations. Nowadays, there are different objective methods to estimate speech quality. Parametric models consider different factors, such as network parameters, acoustic characteristics, communication equipment, among others. The most representative parametric models for telephone service are described in ITU-T Rec. G.107 and G.107.1, mostly known as E-model and WB E-model, respectively. However, they do not consider wireless network parameters as inputs. In this context, this research proposes a speech quality parametric model (SQPM) based on artificial neural networks that considers both wireless network degradation characteristics and the techniques used to improve the transmission quality. For this purpose, a network simulator was built, in which two forward error correction (FEC) codes and four different antenna configurations in a multiple-input-multiple-output (MIMO) system are implemented. To validate the results obtained by the simulator, the ITU-T Rec. P.863 and the WB E-model are used. Experimental results show how different wireless network configurations impact on speech quality. Performance evaluation results demonstrated a high correlation between the proposed SQPM and ITU-T Rec. P.863 results, reaching an PCC and an RMSE of 0.9901 and 0.1492, respectively. Therefore, our proposal intends to be useful for wireless network planning tasks.

Voice quality assessment of phone calls is a relevant task for mobile service providers. In this context, the main objective of this research is to provide a model that improves the performance of ITU‐T Rec. P.563. To accomplish this objective, the proposed model considers two aspects, better response in lossy network and adequate treatment of silences segments into the audio signal. Thus, a function is determined to suppress silences in the speech signal according to the packet loss rate value. Furthermore, the proposed model is implemented on both a voice quality server and a mobile device. Experimental results show that P.563 algorithm performance was really improved by the proposed model, approximating its results to those given by P.862 algorithm, reaching a Pearson correlation coefficient of 0.9957 and a root mean square error of 0.2983. Moreover, subjective test results demonstrated that the proposed model results overcome those obtained by the P.563 algorithm.

Online social networks provide relevant information on users' opinion about different themes. Thus, applications, such as monitoring and recommendation systems (RS) can collect and analyze this data. This paper presents a knowledge-based recommendation system (KBRS), which includes an emotional health monitoring system to detect users with potential psychological disturbances, specifically, depression and stress. Depending on the monitoring results, the KBRS, based on ontologies and sentiment analysis, is activated to send happy, calm, relaxing, or motivational messages to users with psychological disturbances. Also, the solution includes a mechanism to send warning messages to authorized persons, in case a depression disturbance is detected by the monitoring system. The detection of sentences with depressive and stressful content is performed through a convolutional neural network and a bidirectional long short-term memory - recurrent neural networks (RNN); the proposed method reached an accuracy of 0.89 and 0.90 to detect depressed and stressed users, respectively. Experimental results show that the proposed KBRS reached a rating of 94% of very satisfied users, as opposed to 69% reached by a RS without the use of neither a sentiment metric nor ontologies. Additionally, subjective test results demonstrated that the proposed solution consumes low memory, processing, and energy from current mobile electronic devices.

Nowadays, one of the most outstanding new urban transport model is the ridesharing service, in which two or more users share a ride. This transport model reduces costs and the number of circulating vehicles, improving user mobility. In the ridesharing service, the users' quality is a tangible novel evaluation parameter. Consequently, this study treats the use of quality of experience (QoE) in the ridesharing service context, proposing a recommendation system (RS) for ridesharing services (RsRS), which considers user profile information extracted from online social networks (OSN) and user preferences. Thus, the main objective of the proposed RsRS is to improve users' QoE. To this end, the users' profile for the ridesharing service is built based on OSN data, which includes group of users with similar characteristics in the same trip, thus avoiding users with opposite preferences. First, subjective tests are carried out to obtain information on users' preferences and the results are analyzed via machine learning algorithms to determine the various user profiles. The experimental results demonstrate that the random forest algorithm has the best performance, considering OSN data and explicit preferences saved in the proposed solution and only OSN data, for average F-measures of 0.92 and 0.91, respectively. Additionally, a ranking containing a list of recommended users to share a ride is determined using a similarity function, and the results demonstrate that 94.2% of assessors agree with the proposed recommendations. Furthermore, the RsRS has a modular configuration and its integration with a real ridesharing service providers is also discussed.

Nowadays, there are several telephone services based on IP networks. However, the networks can present many disturbances, such as packet loss rate (PLR), which is one of the most impairing network factors. An impaired speech communication affects the users' quality of experience; hence, the assessment of speech quality is relevant to the telephone operators. Therefore, the determination of a methodology to predict a speech quality with a higher accuracy in telephone services is relevant. In this context, this letter introduces a novel nonintrusive speech quality classifier (SQC) model based on deep belief networks (DBN), in which the support vector machine with radial basis function kernel is the classifier applied in DBN, in order to identify four speech quality classes. A speech database was built, based on unimpaired speech files of public databases, in which different PLR models and values are applied, and a standardized intrusive method is used to calculate the index quality of each file. Results show that SQC largely overcomes the results obtained by ITU-T Recommendation P.563. Also, subjective tests are performed to validate the SQC performance, and it reached an accuracy of 95% on speech quality classification. Furthermore, a solution architecture is introduced, demonstrating the usefulness and flexibility of the proposed SQC.

This article analyzes Brazilian Consumers' Sentiments in a specific domain using a system, SentiMeter-Br. A Portuguese dictionary focused in a specific field of study was built, in which tenses and negative words are treated in a different way to measure the polarity, the strength of positive or negative sentiment, in short texts extracted from Twitter. For the Portuguese Dictionary performance validation, the results are compared with the SentiStrength tool and are evaluated by three Specialists in the field of study; each one analyzed 2000 texts captured from Twitter. Comparing the efficiency of the SentiMeter-Br and the SentiStrength against the Specialists' opinion, a Pearson correlation factor of 0.89 and 0.75 was reached, respectively, proving that the metric used in the Sentimeter-Br is better than the one used in the SentiStrength. The polarity of the short texts were also tested through machine learning, with correctly classified instances of 71.79% by Sequential Minimal Optimization algorithm and F-Measure of 0.87 for positive and 0.91 for negative phrases. Another contribution is a Twitter and Facebook search framework that extracts online tweets and Facebook posts, the latter with geographic location, gender and birth date of the user who posted the comments, and can be accessed by mobile phones.