Mental health is essential to overall well-being, influencing physical, psychological, and social health. Despite advancements in healthcare, diagnosing and treating mental health disorders remains challenging. Delayed recognition of symptoms often leads to worsening conditions, emphasizing the need for real-time mental health monitoring solutions. This paper presents an AI-driven mental health monitoring system that continuously assesses emotional well-being. The system integrates a MAX30102 sensor and an ESP32 microcontroller to measure heart rate and oxygen saturation, transmitting real-time data via the Blynk API for continuous analysis. Additionally, Apple Health is utilized to track sleep duration and physical activity, providing a comprehensive evaluation of mental and physical health. By leveraging IoT-based data collection and AI-driven analytics, the system detects stress levels and emotional imbalances through key physiological indicators, including heart rate variability, sleep patterns, oxygen saturation, and physical activity. A random forest regression model achieves a coefficient of determination score of 0.97 and a mean absolute error of 0.17, demonstrating high accuracy in stress prediction. This technology-driven approach enhances mental health monitoring, enabling early detection, personalized intervention, and improved well-being through real-time insights and proactive support.
With the exponential growth of edge devices, the cloud edge continuum provides a natural evolution to the centralised cloud architecture to overcome the bottlenecks created by the growing data that devices generate. Resource constrained edge devices need to be able to offload computational tasks to the cloud edge continuum. Providing resources located at the edge, close to resource-constrained devices, allows devices to offload on demand potentially complex functions with low latency and response time requirements. The COGNIT framework introduces the novel concept of function as a services (FaaS) at the edge and novel AI techniques for cloud-edge management. In this paper, we show how the novel edge FaaS model can be used to offload critical security functions from edge devices and enable them to protect themselves even though they don’t have the resources for such protection. The paper’s main contribution is to show how the edge FaaS model enables to design multi-layer protection models between edge devices and the cloud-edge continuum AI-based orchestrator. In this model the edge device provides a first layer of defense using application knowledge to protect itself whereas the AI-based orchestrator provides a second layer of defense that is more generic because it does not know much about the edge application. The layered protection model is illustrated and validated on a cybersecurity case study where AI-based anomaly detection is deployed at the edge to secure mobile devices and detect anomalies as early and quickly as possible. This second contribution of the paper shows how continuous security anomaly detection can be designed as multiple functions that are triggered by monitored events to provide continuous detection at the edge for all events.
Access logic algorithms govern how occupants gain entry to rooms in the building access control industry by validating credentials presented to electronic locks. The security of these systems depends on the policies used to design the underlying access logic. Quantifying and ranking the security risks associated with different access control policies helps building planners and architects make informed decisions to enhance security. This paper conducts a thorough security risk assessment for building access control systems, following NIST guidelines. It presents a structured method to calculate and evaluate security risks across four widely recognized access control policies - Discretionary Access Control (DAC), Mandatory Access Control (MAC), Rule-Based Access Control (RuBAC), and Attribute-Based Access Control (ABAC). Using a student dormitory model, the study introduces a Risk Ranking Value metric based on features such as room criticality, usage frequency, and building occupancy. The study ranks the aforementioned access control policies using the Risk Ranking Value metric and provides insights into their suitability for various application scenarios.
The rapid proliferation of Internet of Things (IoT) devices and latency-sensitive applications in 5G and beyond networks requires efficient task placement strategies to optimize system performance. Deep Reinforcement Learning (DRL) has emerged as a promising approach to address this challenge, dynamically adapting placement decisions based on real-time network conditions. However, existing DRL-based solutions often neglect the computational overhead introduced by the agent itself, overestimating their effectiveness and limiting their practical applicability. This paper addresses this issue by proposing a novel DRL-based task placement framework that explicitly incorporates the agent’s decision-making time and energy consumption into its optimization objectives. The results show that when DRL computational costs are considered, DRL-based strategies experience 8.43% and 10.45% reductions in average reward, 14.67% and 17.25% increases in response time, and 5.77% and 10.54% higher missed deadline rates compared to when these costs are neglected. Furthermore, the results show that the performance achieved is comparable to that of the Random or Round Robin approaches. This is even more aggravating when analyzing the energy consumed as DRL approaches consume orders of magnitude more energy than Random or Round Robin (ranging from 1,569,336% to 6,849,025%). These results highlight the need to account for the time and energy consumed by the scheduler when making decisions, as the smarter decisions are outweighed by the time the scheduler takes to make them, leading to a reduced overall improvement.
Long Range Wide-Area Network (LoRaWAN) is a promising Internet of Things (IoT) technology, and one of its operational parameters that influences reliable transmission is the Spreading Factor (SF), as this parameter is directly related to receiver sensitivity and packet transmission time. Among the various sectors that can benefit from LoRaWAN, one notable example is Electric Power System (EPS). In this context, this work develops a Spreading Factor Allocation (SFA) scheme, named Reliable - Spreading Factor Allocation (R-SFA), and compares it through simulations with three related solutions: Initial Spreading Factor Allocation (I-SFA), Adaptive Data Rate (ADR), and I-SFA+ADR. The simulated test scenarios execute Advanced Metering Infrastructure (AMI) applications with 200 to 1000 Smart Meters (SMs) distributed over an area of $56.25 \mathrm{~km}^{2}$. The results show that R-SFA is more efficient than alternative schemes, as it reduces the number of Data Aggregation Points (DAPs) by up to $90 \%$, guarantees the reliability rate required by the tested applications, obtains an average Packet Delivery Ratio (PDR) up to $\mathbf{1 4. 4 1 \%}$ higher, and successfully receives packets from SMs located 2272.43 meters away.
Wireless sensor networks (WSNs) are essential in applications such as environmental monitoring and smart agriculture, where they collect and transmit data from remote locations. A key challenge is energy efficiency, especially for battery-powered applications, as it affects network lifetime and maintenance costs. To address this, the design and implementation of a wireless sensor network architecture that combines ESP-NOW and LoRa protocols to create a low-power IoT solution is presented. This hybrid approach leverages the energy efficiency and simplicity of ESP-NOW for short-range local communication between nodes while using LoRa for long-range transmission to the back-end infrastructure (Gateway). The results demonstrate that the proposed architecture offers an efficient and flexible solution compared to traditional architectures based solely on LoRa or WiFi, thanks to the optimized integration of ESP-NOW for the local network and LoRa for long-range transmissions.
Diabetes is a serious chronic condition affecting over 9% of the global adult population. Poor management of the disease can lead to severe complications such as neuropathy and retinopathy. In recent years, with the emergence of technological means, several schemes have been proposed to foster more efficient disease management. However, many existing solutions lack real-time forecasting or adaptive personalization. In addressing the current limitations, this paper proposes a lightweight, AI-driven mobile platform that integrates glucose level prediction and personalized recommendation generation. The system employs third-order polynomial regression trained on real-world glucose data from the Ohio Type 1 Diabetes dataset to forecast short-term glucose trends based on historical user inputs and uses a large language model to generate personalized recommendations. In urgent cases, the system can trigger location-based alerts to notify nearby users or emergency contacts. The proposed regression model achieved an average mean absolute error of $\mathbf{5. 0 8}$ milligrams per deciliter, a root mean square error of 8.21 milligrams per deciliter, and a coefficient of determination of $\mathbf{0. 9 7 6}$. In addition, qualitative testing showed that the language model generated recommendations that aligned with specific user profiles. Overall, the results demonstrated in this paper highlight the system’s potential to support personalized diabetes self-management through a unified and user-friendly mobile platform.
The growing adoption of eXtended Reality (XR) mostly relies on 360-degree video streaming to deliver immersive Virtual Reality (VR) experiences. However, end-to-end encryption of these streams over 5 G networks poses challenges for Mobile Network Operators (MNOs) in assessing user Quality of Experience (QoE), requiring a clear mapping between network-level Quality of Service (QoS) and user-perceived quality. To address this, we propose ASSESSOR 360°, a Mobile Edge Computing (MEC)-based framework designed for QoE estimation in 360-degree VR video streaming. Unlike previous proposed frameworks limited to traditional single-user video streaming, ASSESSOR 360° supports multiple concurrent users, e.g., indoor static, outdoor static, pedestrian, and high mobility (inside vehicles), watching identical or distinct VR videos, enabling analysis of location and mobility impacts on QoE. The ASSESSOR 360° also uses the MEC setup to bring content near to user equipment which helps to minimize the bandwith utilization and latency and it also helps to provide real time experience. By Using real 5G network traces in an emulated environment, we evaluated the framework with encrypted TCP and QUIC traffic. ASSESSOR 360° extracts QoS features at the edge and QoE features from the user side (user logs), providing accurate QoE estimate. This framework equips MNOs with a robust tool to estimate QoE and also investigate QoS-QoE relationships, ensuring high-quality VR video experiences for users under diverse conditions. The framework performs comparative analysis of TCP (HTTP/1.1 persistent) and QUIC (HTTPS persistent) protocols across static and mobility scenarios, revealing QUIC’s superior startup performance but higher stall durations during mobility compared to TCP. Experimental results demonstrate TCP’s stability advantages for sustained 360° video streaming despite its slower connection establishment. ASSESSOR $\mathbf{3 6 0}^{\circ}$ ultimately provides a comprehensive toolkit for evaluating both technical QoS parameters and perceptual QoE metrics in encrypted XR streaming environments.
The accurate diagnosis of psoriasis, melanoma, and eczema is essential for effective treatment. Traditional machine learning-based diagnostic methods faces hurdles in providing good screening performance for these diseases. The deep learning models, like CNNs, Long Short Term Memory (LSTM) networks, coupled with image processing schemes, such as K-means clustering and histogram equalization, presents some improvement in this regard. But still such schemes are limited to accurately represent the morphology, texture, and pliability of the skin in complicated cases. This paper presents a detailed discussion on the pros and cons of different state-of-the-art diagnostic methods which are developed to screen skin diseases. Furthermore, the paper shed light on a new vision-based tactile sensing (VBTS) technology, particularly the DIGIT sensor, that presents good alternative to screening skin diseases alongside dermoscopy, especially in rural areas. Moreover, this paper also presents the results that reveal the importance of the inclusion of VBTS in future developments of dermatological diagnosis systems.
Software-Defined Vehicles (SDVs) is a paradigm of the wider internet of autonomous vehicles where its features and functions are enabled and managed through software, allowing for continuous updates and enhancements throughout its lifecycle. Enabling this technology relies on secure and efficient authentication mechanisms to ensure trusted communication in Vehicle-to-Everything (V2X) networks, over-the-air (OTA) updates, and intra-vehicle Electronic Control Unit (ECU) communication. As SDVs progress to support complex applications such as autonomous driving, platooning, and edge-assisted vehicular computing, selecting an optimal authentication scheme becomes critical for maintaining security without compromising real-time performance. In this work, we present a comparative analysis of state-of-the-art cryptographic authentication techniques and evaluate the impact on authentication latency, energy consumption and computational overhead. This work highlights the effectiveness of different authentication strategies and provides insights into the optimal selection of cryptographic primitives for SDV applications, balancing security and power consumption. Index Terms-Authentication Protocols, Energy profiling, Software-defined Vehicles
Managing cloud storage efficiently is vital in the Big Data era, where changing access patterns affect cost and performance. This study proposes a machine learning framework to predict object access frequency, enabling dynamic tier allocation based on user preferences for cost and latency. Using a fixed $4 \times 4$ week segmentation, the model was trained and evaluated with real Dropbox data across two Points of Presence (PoP-1 and PoP-2). Various classification algorithms, including Support Vector Machines and Logistic Regression, achieved accuracies of up to $79 \%$, with cost reductions of up to $\mathbf{4. 9 5 \%}$ in PoP-1 and $\mathbf{3. 9 \%}$ in PoP-2 compared to traditional online methods. Techniques such as SMOTE were applied to address class imbalance, ensuring prediction reliability. Simulations highlighted the importance of minimizing latency to maintain service quality. The results demonstrate the effectiveness of machine learning in optimizing cloud storage, providing significant economic benefits without compromising service quality.
As the development of 6 G networks accelerates, integrating Artificial Intelligence (AI) becomes essential for achieving ultra-fast data rates, low latency, and intelligent resource management. However, implementing AI at the edge introduces significant challenges, particularly under privacypreserving frameworks like Federated Learning (FL). While FL mitigates data privacy concerns by keeping data on user devices, it also introduces high computational demands on the user side and complicates model training due to the unbalanced distribution of the data. This paper presents a Predictive Performance Modeling (PPM) approach that utilizes meta-learning to model key FL training outcomes, such as model accuracy, execution time, and the required number of global epochs. By analyzing historical FL training data across various datasets and hyperparameters, our approach achieves promising predictive accuracy, with Coefficient of Determination ($R^{2}$) scores of 0.7, $\mathbf{0. 8 5}$, and $\mathbf{0 . 5 2}$ for accuracy metric, execution time, and number of epochs, respectively. These results highlight the potential of PPM to enhance the efficiency and scalability of FL, enabling more practical AI deployment in the resource-constrained, privacysensitive landscape of 6 G networks.
Network Digital Twins (NDTs) aim to provide high-fidelity virtual representations of networks, replicating their architecture, components, and real-time dynamics. They enable organizations to analyze, optimize, and predict network performance and behavior under diverse scenarios. As telecommunication networks become increasingly complex – driven by advancements such as edge computing, 5G/6G, and software-defined networking – the need for advanced data analytics techniques like Deep Learning (DL) becomes more pressing. Despite their accuracy and efficiency, DL-based NDTs face limited adoption due to their lack of explainability. To address this challenge, our study builds on previous work that proposed a Graph Neural Network (GNN)-based NDT model by integrating an eXplainable AI (XAI) framework to enhance its interpretability. Our findings reveal the key input-output features that most influence the GNN-based NDT model’s predictions, providing deeper insights into its decision-making process. Additionally, we present practical recommendations for improving network operations, with a particular emphasis on sustainability.
Collaborative cloud–fog–mist computing has been systematically developed to optimize the trade-off between efficiency and data privacy in executing complex applications with diverse security requirements. In these environments, workloads often exhibit a complex structure, comprising jobs with varying levels of security demands. Consequently, it is imperative to implement security-aware scheduling schemes to ensure the correct execution of these applications. This research investigates security-aware scheduling policies, focusing on appropriate algorithms for mixed workloads, which include both simple single-task jobs and Bags of Linear-Workflows (BoLWs). Mist resources are allocated to process simple jobs with the highest security requirements, designated as mist jobs. Simple jobs classified as fog jobs, which have intermediate security levels, can be executed on either fog or mist resources. BoLWs, with the lowest security level, are suitable for execution on cloud or fog resources. Multi-criteria scheduling algorithms are employed to prioritize mist and fog jobs, ensuring satisfactory performance. Extensive simulation experiments are conducted to evaluate the effectiveness of these algorithms across varying levels of system utilization, security requirements, and variability in job/task service demands. The results demonstrate that the performance of the scheduling schemes is contingent upon the level of variability in service demands and the system load.
Despite current anti-phishing measures, businesses still fall for phishing attacks, this issue stems from modern phishing attacks leveraging artificial intelligence. To address this issue, this research proposes the use of an automated pipeline to generate spear phishing emails for use by internal security professionals in practical spear phishing exercises. The proposed pipeline is automated, capable of using real employee data, is fully customisable with modular components, accepts any input data with fine-tuneable outputs, and is portable, as it can be accessed through the web. The pipeline was well received by businesses with generally positive feedback on the effectiveness of the pipeline and potential use cases. The generated emails were of good quality and of 60 responses, on average, the participants rated the emails a score of $4.11 / 5$ or “Very Accurate” and “Likely” to interact with the content of the emails, with an average score of $3.02 / 5$. The pipeline took an average of $\mathbf{1 7. 5}$ seconds per email. Using this number, it would take a business of 100 employees 29.17 minutes or a larger business with 1000 employees 4.86 hours. This level of speed and effectiveness showed that the pipeline could serve as an effective training tool for businesses.
Deep neural networks (DNNs) are increasingly used for real-time image processing, augmented reality, mobile robotics, and autonomous driving. However, IoT devices often have limited processing power and energy and must contend with varying network conditions. Meanwhile, edge computing reduces delays by processing data closer to its source; however, edge servers can become saturated or overloaded under heavy workloads, potentially leading to degraded performance and an inability to meet service quality requirements. This paper introduces a new Deep Neural Network Partitioning Strategy (DNNPaS) for Optimizing End-to-End Latency in Dynamic Mobile Edge Computing. DNNPaS uses an external machine learning model to predict the latency of each DNN layer and data transmission. This external machine learning model gets the client and edge device metrics along with network conditions to predict the optimal partition point of the DNN model. This prediction minimizes end-to-end latency by effectively balancing the computational workload between on-device processing and offloading to the edge server. Experimental results on both sequential and non-sequential DNN architectures demonstrate that our approach reduces end-to-end latency by up to $2 \times$ compared to server-only processing, achieving an overall latency reduction of $\mathbf{3 0 - 5 0 \%}$ under varying load conditions while efficiently utilizing available resources. Our method dynamically adjusts to diverse network environments and server conditions, offering a robust solution for deploying DNNs on resource-constrained devices with improved overall performance.
Regulatory sandboxes have become a vital tool in the European Union for fostering innovation while ensuring regulatory compliance. Although widely adopted across various sectors, there remains a lack of operational procedures specifically tailored to define and test cybersecurity requirements in such protected environments, for example for Internet of Things (IoT) products, which are increasingly critical yet vulnerable. This paper addresses this gap by proposing a multi-stage framework to define and develop sandbox testing requirements aligned with the Cyber Resilience Act (CRA). Our approach incorporates different levels of granularity to map cybersecurity requirements, ensuring adaptability to different product types and organizational contexts. We validate the framework through a Proof-of-Concept on IoT devices, highlighting challenges such as misaligned standards and correlations between CRA provisions and the Italian National Cybersecurity Perimeter Law. The proposed framework aims to streamline regulator-organization interactions and enhance legal certainty in cyber resilience testing within regulatory sandboxes.
The application of edge caching for media processing in AI inference latency reduction remains an unexplored area of edge computing research. This paper investigates how strategically placed edge caches can speed up AI model inference on immutable images because these images do not change after creation thus simplifying cache invalidation processes. The research focuses on model optimization techniques but we examine data access pattern optimization through proximity-based edge caching as a complementary approach. The proposed architecture undergoes simulation evaluation at different cache hit rates ranging from 0% (cold) to 95% (fully warmed) to measure end-to-end latency and throughput improvements. The results show that highly-warmed edge caches decrease inference processing time by 60.22% compared to origin-based processing while achieving more than 150% throughput improvement. The research expands current knowledge about edge AI optimization by demonstrating data locality matters for social media content moderation and real-time visual analytics and large-scale media classification systems because fast processing speeds directly affect user experience and operational efficiency.
Automated inspection of metal surface and edge defects is important. In this study, we have designed and developed an optimal setting to detect surface and edge defects of metal plates. We conducted four consecutive experiments where we applied several artificial intelligence algorithms, ResNet18, ResNet50, ResNet152, Xception, ShuffleNet, GoogleNet, and multiple YOLO algorithms (YOLO5, YOLO8n, YOLO8s) to inspect the optical data set obtained. In addition to the set of real defects, we have trained and tested our models with synthetic data. We have compared the performance metrics of the models based on two real and augmented data sets. On different sets, we found that GoogleNet and ShuffleNet reached 100% accuracy during testing. We also reported the results of the YOLO models that were trained and tested on the data sets. In the future, we are planning to develop a custom network and perform an ablation study.
Artificial Intelligence delivers automated analysis, cognition, and decision making based on several approaches, such as knowledge representation, multi-agent systems, planning, and machine learning. AI-based applications have increased requirements in terms of access and processing of data as well as computing capacity. For example, in their training/learning phase, AI algorithms rely on a significant amount of storage and computational resources so that huge amounts of training data can be fed into the algorithms. Such processes have been recently shown to have massive carbon footprint. The computing continuum, ranging from centralised cloud data centres through edge and fog nodes to IoT devices, is on a growth trajectory to play a crucial role in the implementation of advanced data-driven applications, including emerging AI- and machine-learning-based applications. With its distributed and federated compute model, the computing continuum will help to address the pressing challenges in exploiting data generated at the network edge and by IoT devices. Currently, the development, deployment and operations of data-driven applications on the computing continuum requires a significant amount of effort and lacks the effectiveness that can yield significant gains for both application and infrastructure owners. In this position paper, the challenges in the orchestration of the computing continuum are described, presenting also the principles of a proposed platform and a set of integrated development, deployment and operations tools that will pave the way for more productive delivery of highquality data-driven applications on the computing continuum.
