Self-attention mechanism (SAM) has demonstrated strong analytical power in fields like computer vision, natural language processing, and genomics, but its use within the domain of cybersecurity is exploratory. A growing body of research highlights SAM’s applications in critical cybersecurity areas, including intrusion detection, malware analysis, adversarial defense, privacy protection, and data security. Despite these advancements, the current literature largely focus on isolated aspects of SAM’s performance, such as model robustness and data confidentiality, while overlooking comprehensive evaluations of its inherent limitations and vulnerabilities. This paper reviews impactful studies employing Security Assurance Metrics (SAM) within security contexts, critically analyzing emerging research trends, relevant methodologies, and case studies. Lastly, it addresses key challenges and proposes directions for future research aimed at advancing SAM’s contributions to the evolving landscape of cybersecurity.
Efficient task execution is critical in mission-driven Edge-Computing (EC)-based IoT applications such as UAV-assisted weather nowcasting. Although UAV fleets improve the resolution and responsiveness of short-term forecasts, their limited on-board computation and battery life demand a cooperative task offloading strategy. Volunteer vehicular edge networks, leveraging idle onboard units of nearby vehicles, offer scalable, low-cost resources. To incentivize task offloading from UAVs to self-interested vehicles, in this work, we introduce a distributed Dutch auction with second-price payments. The proposed scheme is suited for efficient and time-sensitive allocations. Furthermore, to overcome the challenge of non-trivial task valuation, we embed lightweight Reinforcement Learning (RL) agents that dynamically learn valuation policies. Conducted experiments demonstrate that our approach achieves up to -93% in information revelation compared to sealed-bid auctions, up to -50% the communication overhead of the best baseline, and linear computation complexity, while preserving high social welfare, which is further improved by the integration of RL.
Communication barriers faced by deaf and mute individuals have a significant impact on their professional, academic, and social engagements. With the advancement of machine learning technologies, there is an increasing need for effective autonomous sign language translation systems. However, existing research has often neglected user comfort and accessibility, leading to a limited presence of practical autonomous sign language translation devices in the market. This paper presents the leading design and application difficulties involved with translating lab-tested sign language translation systems into real products and offers potential solutions. Our proposed system architecture is modeled using SysML in yEd, capturing the overall system’s functionality and interaction. Additionally, we test quantized MobileNet CNN models and the OpenFace toolkit and benchmark their performance to popular computer vision solutions in the context of portable sign language translation devices. We aim to create a blueprint for complete portable end-to-end sign language translation systems in the future that can transition into real-life products, as a product that successfully addresses user comfort, pricing, and translation accuracy could substantially improve the lives of millions of deaf and mute users around the world.
Digital twin, a popular tool taking advantage of recent advancements on the Internet of Things (IoT), Machine Learning (ML), and Big Data, has rapidly received considerable attention in several fields with applications extending from smart city management to advanced VR/AR systems. We review the detailed investigations into the current research landscape related to digital twins, examining their use across these different fields, and comparing their applications, challenges, and potentials. In the realm of smart cities, creating a digital replica of the city infrastructure could significantly enhance urban planning, city management, and disaster response. Applying the digital twin technology within the smart city would surely bring its unique challenges and opportunities and meet the increasing requirements of city development. Additionally, as a platform for modeling and simulations in the manufacturing industry, digital twin creates virtual copies of these physical entities. By doing so, the digital twin application collects and analyzes unexpected data in manufacturing scenarios which eventually provides better decision making and increases productivity. In the VR/AR industry, the potential of digital twins reveals that this technology could be used for remote working environments, predictive health diagnostics, and personalized treatment plans for the healthcare domain. Although promising, digital twin technology also faces considerable challenges such as the complexities of data communication and accumulation, the dearth of data for training ML models, the need for massive processing power to support high fidelity twins, the high demand for interdisciplinary collaboration, and the absence of standardized development methodologies and validation measures. This paper, however, focuses on the positive aspects and benefits of digital twin technology.
The rapid adoption of electric vehicles (EVs) raises critical challenges in managing charging stations, requiring efficient resource allocation to balance demand, optimize energy use, and maintain grid stability. This paper proposes FL-PSO, a hybrid strategy combining Federated Learning (FL) and Particle Swarm Optimization (PSO) for resource management in fog-cloud computing environments. FL-PSO is applied to EV charging stations and benchmarked against standard PSO, Genetic Algorithm (GA), and Cat Swarm Optimization (CSO). Evaluation is conducted using Quality of Service (QoS) metrics, including latency, energy consumption, and load balancing. Results show that FL-PSO achieves a fitness value of 0.1489, reduced energy consumption (0.2644 W), and improved load balancing (4.29), consistently outperforming baseline methods under both fixed weights and optimal QoS trade-offs. By leveraging the synergy between FL and PSO, FL-PSO provides a scalable and intelligent solution for distributed fog-cloud systems supporting smart EV charging infrastructures.
The Internet of Things (IoT) connects billions of devices across domains such as transportation, healthcare, agriculture, and energy-creating highly dynamic, distributed, and heterogeneous environments. These characteristics pose significant challenges for control, coordination, scalability, and adaptability. In response, Multi-Agent Deep Reinforcement Learning (MADRL) has emerged as a promising paradigm by combining the decision-making intelligence of reinforcement learning with the collaborative capabilities of multi-agent systems. To leverage MADRL in building intelligent, resilient, and adaptive IoT systems, this review systematically explores the application of MADRL in IoT, categorizing contributions by application domains, learning architectures, and coordination strategies. We analyze how MADRL enables scalable resource allocation, routing optimization, energy efficiency, and fault detection in complex IoT ecosystems. Furthermore, we highlight key challenges-including scalability, non-stationarity, partial observability, and communication overhead-and discuss emerging solutions such as mean-field approximation, belief-state tracking, and federated MADRL.
This study investigates the use of machine learning to predict budget compliance in public construction projects using Portuguese procurement data. Building on prior work, we reproduce and enhance a Random Forest classifier by applying rigorous data preprocessing, SMOTE for class balancing, and systematic hyperparameter tuning. A refined classification target based on the Effective Price to Initial Price ratio allows the model to distinguish over-budget, compliant, and under-budget contracts. We further improve model interpretability through a multi-step feature selection process and unsupervised analysis, including UMAP and Kernel PCA. The resulting model achieves an accuracy of 89% and a weighted F1-score of 0.88-out-performing the original study. These results demonstrate that well-tuned classical models can deliver high performance while maintaining transparency and computational efficiency, making them suitable for supporting early-stage decisions in public procurement workflows.
Millions of people worldwide live with visual impairments like glaucoma, age-related macular degeneration (AMD), and diabetic retinopathy, conditions that hinder daily activities, facial recognition, and mobility, ultimately reducing independence and quality of life. As these conditions remain incurable, current treatments mainly aim to slow progression rather than restore vision. Most smart glasses follow a one-size-fits-all model, overlooking individual vision loss patterns and disease evolution. To bridge this gap, we developed a personalized prototype of smart glasses that adapt to each patient’s unique visual field using data from Humphrey visual field tests. The system detects the widest non-blind area and dynamically adjusts the display to project the full scene within that region, allowing users to see everything through their remaining functional vision. Our goal was to deliver a highly personalized solution by integrating patient-specific data with advanced simulation techniques for the most accurate and effective visual enhancement. The prototype combines two mini Wi-Fi cameras with a virtual reality (VR) headset to simulate and adjust vision in real time. Early results show promising improvements, offering a personalized and adaptive approach that redefines visual assistance and promotes greater independence and quality of life.
Federated learning enables distributed clients to train a global model while maintaining privacy and control over their data. Although collaboration can substantially improve the learning process, it also introduces vulnerabilities, as not all participants contribute beneficially. Clients may engage in detrimental activities, such as data poisoning, that compromise the integrity of the global model. Additionally, in a realistic scenario, the quality of the data possessed by these clients is highly heterogeneous, which influences the model training performance. Moreover, malicious clients (aka free riders) intend to obtain the global model without making a real contribution to the training process. Hence, a reliable and fair evaluation of the client contribution is essential to promote diverse client engagement, improve robustness, and address the free-rider problem. This paper proposes a reputation-aware contribution evaluation approach (RBFL) that provides adversarial robustness by tracking reputation over multiple training rounds to ensure that clients consistently contribute positively. We employ CosineGradient as the utility function and Truncated Monte Carlo (TMC) Shapley as the data valuation function. Empirical evaluation demonstrates the effectiveness of our approach in a fair evaluation of clients' contributions and effective identification of adversarial clients while maintaining a model accuracy of 92% with adversarial robustness.
With the growing adoption of Artificial Intelligence (AI) in sensitive sectors such as healthcare and finance, protecting user privacy during data processing has become paramount. One promising approach is Fully Homomorphic Encryption (FHE), which offers a viable solution by allowing computations to be performed directly on encrypted data, thus safeguarding sensitive information. In this study, we investigate the practical application of the Cheon Kim Kim Song (CKKS) FHE scheme to perform inference with various machine learning models for heart disease detection. We evaluated five models: Logistic Regression, Support Vector Machine, Decision Tree, Random Forest, and a simple Neural Network, across multiple heart disease datasets. Our analysis compares their performance on both standard (plain-text) and encrypted data, using metrics including accuracy, precision, recall, and F1-score. Results demonstrate that encrypted models deliver predictive accuracy comparable to their standard counterparts, confirming the viability of privacy-preserving inference with FHE despite the expected increase in computational time. Furthermore, our findings highlight up to 100% consistency between the predictions made on encrypted and plain-text inputs.
GraphQL provides developers with an alternative to REST API systems by allowing for only the exact data necessary to be grabbed for a client using a singular endpoint. This technology, while offering numerous benefits, is still new in the development sector and requires further resource and dedication to find and implement complementary resources. In this paper, we continue the development of a multilevel architecture based on a GraphQL implementation using Hasura. The architecture focuses on a React application and a Hasura instance from which we ensure the smooth addition of data visualization, governance, and quality control. We review the process to final and set-up the technologies used to ensure these three components and the pros and cons of these systems as well as the entire architecture.
Food safety and authenticity are critical global concerns, increasingly challenged by the complexity of modern supply chains and the rise of sophisticated fraud techniques. In this paper, we propose a novel approach for detecting honey adulteration by combining hyperspectral imaging with a generative AI framework based on deep autoencoders. Our method leverages the autoencoder’s ability to model the intrinsic spectral patterns of pure honey, enabling the detection of subtle anomalies introduced by adulterants. Through a rigorous binary classification task, our system achieves a high F1-score, demonstrating both accuracy and robustness. In addition, we provide an indepth analysis of reconstruction quality and error distributions, highlighting the capacity of the model for interpretable and reliable fraud detection. The proposed methodology offers a scalable and data-efficient solution with broad applicability to food authentication, paving the way for future deployment in real-time quality assurance systems.
Graph Neural Networks (GNNs) are increasingly applied in domains such as drug discovery, medical diagnostics, and community detection. Despite their powerful predictive capabilities, GNNs remain vulnerable to privacy threats, particularly in sensitive contexts like healthcare and finance where data confidentiality is paramount. This study targets a pressing security concern in GNNs by addressing the risk of data exposure through Membership Inference Attacks (MIA). We propose a novel, model agnostic defense framework based on Diffusion based Posterior Smoothing (DPS), which enhances the privacy of node level predictions by perturbing class probabilities through a learned diffusion denoising process. Unlike prior methods, DPS maintains the predictive utility of the GNN while significantly reducing the information leakage exploited by MIA adversaries. Empirical evaluations across diverse datasets and GNN architectures confirm that DPS provides robust protection against inference attacks with minimal compromise to node classification accuracy.
Diabetic macular edema (DME) and age-related macular degeneration (AMD) are major causes of vision impairment and blindness. While many classification applications for these diseases achieve high performance, they often overlook the crucial aspect of dataset and image quality, leading to potential erroneous predictions. This study emphasizes the importance of data quality in medical image classification, specifically for retinal imaging. We propose an Optical Coherence Tomography (OCT) image quality evaluation model using the pre-trained ARNIQA (leArning distoRtion maNifold for Image Quality Assessment) model to accurately identify retinal diseases autonomously. Our methodology includes a three-class classification system utilizing two Convolutional Neural Network (CNN) models, ResNetS0 and Xception, applied to three datasets: the original dataset, a subset of high-quality images, and a subset of low-quality images. Using a Tunisian OCT dataset of 2887 images, we demonstrate the efficacy of our approach, achieving 100% accuracy with high-quality images.
The exponential rise in security alerts across modern information systems presents a critical challenge for cybersecurity operations, frequently overwhelming analysts and hindering timely incident response. To address this, we propose an automated threat classification framework that leverages RetrievalAugmented Generation (RAG) in combination with Large Language Models (LLMs) to accurately interpret and categorize attack types. By combining real-time data ingestion with contextual retrieval and advanced natural language understanding, the system efficiently interprets alert data and categorizes threats to support incident response and decision-making. Experimental evaluations show that the proposed approach achieves a classification accuracy of 94%, demonstrating strong promise for realworld deployment in security operations centers. These results highlight the potential of integrating RAG-LLM based frameworks to significantly enhance the scalability and effectiveness of cybersecurity defenses.
Radio Frequency Identification (RFID) technology is increasingly utilized to provide an efficient identification and tracking capabilities across a wide range of domains. In the medical field, RFID technology enable patient monitoring and equipment management. However, the constrained resources of RFID tags make them vulnerable to security threats, raising concerns about data integrity, secure access, and patient privacy. This paper proposes a novel group-based RFID authentication protocol that introduces a threshold-based mechanism that enables group authentication even when some tags fail to respond. We conduct both formal and informal security analyses using tamarin prover to evaluate the proposed protocol. In addition, we analyze its performance using a thresholdbased availability model. The proposed protocol achieves improved resilience, privacy, and fault tolerance, making it well suited for sensitive Internet of Things (IoT) environments.
In evaluating semantic similarity, word senses are crucial, and word senses vary with context. In the literature, existing word sense disambiguation methods are often expensive, limited, and may require human intervention. Moreover, identifying the correct sense of words can often be subjective even for humans. For these reasons, we propose a fuzzy approach that incorporates all possible contexts and associated senses of words by leveraging the BabelNet encyclopedic dictionary. A previous experiment using the most representative taxonomy-based similarity measures demonstrates that our proposal improves Pearson correlation with human judgment for all selected measures.
We propose VOST, a novel segmentation-bytracking framework for semi-supervised Video Object Segmentation (VOS) that decouples the tracking and segmentation tasks to improve both accuracy and robustness. Our approach employs a state-of-the-art zero-shot tracker (SAMURAI) to generate bounding boxes of the target object in each frame. These cropped regions are then fed into a Vision Transformer (ViT)-based segmentation model trained to segment the object within the bounding box. By isolating the object from distracting background content and similar instances, our model eliminates ambiguity, simplifies the segmentation task, and requires no temporal memory to maintain object consistency. VOST is evaluated using three benchmark datasets, DAVIS16, DAVIS17, and SegTrackV2, achieving state-of-the-art performance. Specifically, VOST reaches an M score of 92.6 on DAVIS16, 88.9 on DAVIS17, and an F-score of 0.929 on SegTrackV2, outperforming all previous methods. Additionally, VOST achieves a real-time inference speed of 20.6 FPS, offering an efficient and scalable solution for practical VOS applications. These results demonstrate the effectiveness of the segmentation-by-tracking paradigm and its potential as a competitive alternative to memory-based approaches.
In this paper, we investigate the application of Genetic Algorithms (GAs) and their parallelization via the Island Model (IMGA) for polygonal image reconstruction. The objective is to approximate a target image using a fixed number of colored triangles, formulating the problem as a high-dimensional optimization challenge. A standard GA is first implemented to assess its effectiveness using various objective functions. We then introduce a parallel IMGA framework, featuring multiple subpopulations evolving independently with periodic migration in a ring topology. Migration parameters are calibrated, revealing optimal settings for preserving diversity and improving convergence. Experimental results indicate that IMGA achieves up to a 50% improvement in reconstruction quality compared to the standard GA, with minimal additional computational cost. Moreover, our approach significantly outperforms the current state of the art for this problem, which relies on a hybrid method combining genetic algorithms and machine learning. These findings highlight the effectiveness and scalability of island models in addressing complex, high-dimensional optimization problems such as image reconstruction.
Quantum machine learning (QML) offers new perspectives for modeling complex sequential data by leveraging the computational advantages of quantum computing. In this work, we explore the application of Quantum Long Short-Term Memory (QLSTM) networks for sequence prediction tasks. By integrating parameterized quantum circuits within the classical LSTM architecture, QLSTM models exploit quantum parallelism and entanglement to learn temporal dependencies in data more efficiently. We evaluate the performance of QLSTM on synthetic and real-world sequence prediction tasks and compare it with its classical counterpart. The results demonstrate that QLSTM models achieve competitive performance with fewer trainable parameters, indicating their potential as lightweight and efficient models for sequence learning and prediction.
