Public key X. 509 certificates play a powerful role in promoting effective electronic identification, but some significant practical issues still affect their scalability. Every time a public key certificate is used, it must be validated by the system or application relying on it for security services, generically called also relying party. The validation involves several processing steps and checks, and it has been measured that many applications (still) perform it incompletely. Furthermore, privacy issues may occur when validating certificates, for example a website visited by a user could be revealed to external parties. We propose TPValCert, an architecture tailoring a trusted proxy to provide privacy-preserving certificate validation service to the relying parties. By exploiting TPValCert, a desktop, IoT, or mobile system that needs to validate a public-key certificate may execute a transaction with a trusted proxy, which performs validation by considering certificate policy parameters, privacy, and validation options received from the client and returns the validation status. Besides reducing complexity on the client, exploiting such trusted validation parties may also bring privacy benefits. To communicate with the clients, we consider the SCVP (Server-based Certificate Validation Protocol) or DVCS (Data Validation and Certification Server) protocols, even though, depending on the context, lighter formats could be considered. Our implementation efforts emphasize the possibility of pursuing a tradeoff between timeliness, privacy, and computational resource usage, via dynamic selection of several configurable options.
This study presents a fine-tuning of the BERTimbau model for Named Entity Recognition (NER) in Electronic Health Records (EHR), focusing on oncology data in Brazilian Portuguese. The proposed approach employs an annotated dataset formatted in the Inside-Outside-Beginning (IOB) scheme, enabling precise extraction of clinical entities. The model was trained using the BERTimbau Base Tokenizer, with hyperparameters optimized through the Hugging Face library. Experimental results demonstrate robust performance, achieving F1scores between 0.83 and 0.85 across disease, procedure, and medication categories. When applied to a dataset of 125,825 new clinical records, the model achieved a correct identification rate exceeding 92%, underscoring its potential for medical decision support. Key challenges include enhancing classification accuracy for less frequent entities and adapting the model to diverse clinical contexts. Future directions involve dataset expansion, data augmentation techniques, exploration of more advanced transformer-based models, and integration into hospital decision support systems to further improve healthcare analytics.
Network Function Virtualization (NFV) provides a flexible mechanism for deploying Virtual Network Functions (VNFs) within Service Function Chains (SFCs), thereby streamlining data transfers between end-users and edge/cloud resources. The distinct requirements for forward and backward traffic—each carrying different types of content—give rise to Hybrid SFCs (HSFCs), which must be carefully designed to address unique deployment and performance concerns. However, achieving robust disaster resilience in HSFC-based NFV environments poses significant challenges, as natural or hardware-induced disruptions within Disaster Zones (DZs) can degrade service quality or even cause outages. This paper describes the Resilient Hybrid Service Function Chain Resource Optimization (R-HSFC-RO) approach to ensure both efficient resource utilization and sustained service delivery under disaster conditions. Our model considers bandwidth consumption, computational resource allocation for VNF execution, VNF instantiation overheads, and end-to-end latency requirements. For resolution, we propose a Mixed-Integer Linear Programming (MILP) model and Constraint Programming (CP), thereby enabling optimal solutions. Simulation results demonstrate that R-HSFC-RO reduces total costs by up to 50%, enhances disaster resilience, and maintains high operational efficiency.
Multiomics is an emerging biological analysis approach in which the datasets come from multiple “omics”, such as genomics, epigenomics, transcriptomics, proteomics, metabolomics, and microbiomics. Nowadays, the convergence of Deep Learning and multiomics sciences presents an unprecedented opportunity to dissect the intricate interplay of biological processes. Specifically, multiomics data integration, propelled by Deep Learning methodologies, has revolutionised biological research, enabling a more holistic understanding of complex biological systems and disease mechanisms. This paper explores the current landscape of Deep Learning applications in multiomics, highlighting state-of-the-art techniques, emerging research areas, and the challenges that lie ahead. In particular, we delve into the application areas and computational methods that have been considered so far, offering guidance to researchers navigating this intricate field.
Open Cloud (O-Cloud) is the Open Radio Access Network (O-RAN) computing infrastructure spanning edge to cloud sites defined by the O-RAN Alliance to support and coordinate RAN infrastructure management shared among multiple Mobile Network Operators (MNO). In the standard definition, O-Cloud ensures reliable connectivity, active coordination, and efficient distribution of the Radio Access Network (RAN) deployments. Thanks to these properties, O-Cloud can support Mobile Network and Infrastructure operators to achieve fully automated infrastructure management, self-service MNO portals, enhanced security measures, and O-Cloud infrastructure-agnostic management. In this paper, we present TORNADO: TOSCA-enabled Orchestration for RAN in Next-Generation Networks Automating DevOps in O-Cloud, an O-Cloud automation infrastructure designed to ease DevOps deployment and operation phases of RAN network functions for multiple MNOs. Our solution introduces infrastructure-agnostic automation for multi-site, multi-MNO RAN components, offering high-level, secure self-service MNO portals for defining RAN deployments. Performance evaluation of our solution for various RAN network functions demonstrated the capability of TORNADO to automate the deployment in a multi-site, multi-MNO heterogeneous infrastructure.
The increasing reliance on the judiciary to secure access to healthcare in Brazil has created a vast corpus of legal rulings that contain insights into healthcare delivery. However, extracting actionable information from these unstructured texts remains a significant challenge. This study evaluates the performance of state-of-the-art large language models (LLMs) for healthcare-related named entity recognition (NER) in Brazilian judicial decisions. Using our release dataset, LexCare.BR, a manually annotated gold-standard dataset of 1,200 legal rulings, we assessed 10 predefined healthcare entities across multiple LLMs, including both open-source and closedsource models. Our results show that larger models, such as GPT-4o and Llama 3.1, achieved the highest overall F1-scores (0.739 and 0.694, respectively), demonstrating robust capabilities in extracting clinically and policy-relevant entities. However, we observed significant variations in performance across entity types, with higher precision for standardized codes, such as ICD-10, but lower recall for context-dependent categories like dietary supplements. These findings highlight the potential of LLMs to automate the extraction of structured healthcare data from judicial texts, thereby enabling real-time monitoring of healthcare judicialization and informing targeted policy interventions. Data is available at https://github.com/eliasjacob/lexcare.br
IoT botnets have been adopted as the prime infrastructure for a plethora of cybercrime and modern cyberwarfare. Evidently, conventional defence approaches fail to capture the full spectrum of IoT botnet activity by virtue of attackers evading schemes and limited Internet visibility. In this work, we develop a novel macroscopic analysis framework that profiles malware strains through payload signatures gathered from malicious traffic, revealing distinct botnet variants and their infrastructure. Through payload clustering distilled by information retrieval properties and DNS-based infrastructure mapping, we systematically group botnet families, identifying distinct exploitation trends and infrastructure reuse patterns. Our longitudinal study over real pre-captured datasets for a $\mathbf{4}$-year period reveals widespread lack of blacklist coverage with $65.94 \%$ of discovered malicious IPs and $\mathbf{9 8. 9 7 \%}$ of associated domains not yet blacklisted. We pinpoint a growing trend of botnet operators leveraging cloud services such as AWS, OVH, and Linode, hosting their command-and-control (C2) servers on reputable domains to bypass security filters and extend operational longevity. Through demonstrating practical metrics to assess botnet scan volume, vulnerability trends, and infection rates, we stress the need to refine existing defence mechanisms. In parallel, we set solid ground for practical threat hunting and risk profiling for next-generation cybersecurity schemes.
Federated Learning (FL) is increasingly used in healthcare to enable collaborative model training across decentralized medical institutions while preserving patient privacy. Despite its promise, FL faces significant challenges in medical image processing, such as class imbalance and client data heterogeneity, which can lead to biased models and reduced accuracy in detecting rare diseases. To address these issues, we propose a novel approach, called Syndicated Federated Learning (SyndFL), comprising a multi-layer client selection algorithm that emphasizes fair client representation based on dataset size, learning performance, label distribution, and domain-specific features. SyndFL not only prioritizes clients from minority data clusters but also includes adaptive weighting to ensure that rare conditions receive adequate representation, reducing the risk of bias in model aggregation. Experiments conducted on healthcare image datasets demonstrate that SyndFL achieves a $\mathbf{1 0}-\mathbf{2 0 \%}$ improvement in detecting rare conditions compared to standard FL methods, significantly enhancing both model robustness and fairness in clinical decision-making and diagnostics.
Quantum computers pose a threat to parts of cryptography in use today. Fortunately, post-quantum cryptography (PQC) provides replacements believed to be secure against quantum computers. Transitioning away from cryptography vulnerable to quantum algorithms is becoming urgent with the advancement of technology in quantum computing, and the release of standards for PQC algorithms. Hybrid cryptography uses both PQC and classical cryptography in parallel, ensuring security in case either of the two algorithms is compromised. Hybrid cryptography is especially useful during the transition to PQC, when the security of the relatively new algorithms has not been time-tested. We have implemented a hybrid signature scheme in the context of a specific application: Self-Sovereign Identity. There are multiple decisions during the implementation of hybrid signature schemes in this context: the combiner used, how keys are stored, and how the hybrid signatures are encapsulated in relevant web standards. Our implementation uses strong nesting for hybridisation, a specific pair of algorithms rather than a general combination, and encodes both the key and signature as a single entity, rather than as two separate components. In addition to an overview of the relevant concepts, we present arguments for the decisions made, as well as related proposals for standardization.
Smart homes are transforming domestic environments by integrating connected devices and sensors, enabling lighting, temperature, and security automation. While these systems enhance comfort and efficiency, they often rely on predefined settings or manual input due to the absence of adaptive recommendation systems. AI-driven recommendation systems personalize actions by learning from user behavior and environmental data, improving the smart home experience. However, they also introduce cybersecurity risks, particularly data poisoning attacks, where manipulated data disrupts system functionality. This paper exposes and examines vulnerabilities in smart home recommendation systems, categorizing data poisoning attacks and analyzing their impact. Through a literature review and attack vector analysis, we identify key weaknesses and propose mitigation strategies to enhance security. Our goal is to contribute to developing robust smart home technologies that protect user privacy, ensure reliability, and withstand adversarial threats.
Recent advancements in Machine Learning (ML) and MLOps for Industry 5.0 have significantly boosted productivity in manufacturing by enabling predictive maintenance and optimizing industrial workflows. However, implementing ML applications in real-world industrial environments presents several challenges, including limited access to labeled data, stringent privacy requirements, and the decentralized nature of industrial data. An effective solution for distributed learning with unlabeled data is essential to address these issues. In this paper, we introduce FedEdge-Learn, a novel Federated Learning framework tailored for Industry 5.0 applications. It focuses on unsupervised K-means clustering enhanced by globally shared data. Our approach safeguards data privacy other than accelerating the onboarding of new machines by leveraging the globally trained model. We validate our framework using both public datasets and real-world industrial data, demonstrating its effectiveness in real-world scenarios. The results show how, with our framework, the K-means algorithm is effective in federated settings, without a significant performance decrease compared to the centralized case.
Traditional Cloud-based multimedia transcoding approaches struggle with issues like latency and bandwidth limitations. In this work, we experiment with strategies that exploit Edge computing capabilities to get around such constraints. We focus on an on-demand multimedia transcoding service for real-time video streaming on the uplink, leveraging a state-of-the-art Edge Computing service orchestrator to handle its provisioning. We regard this as a use case for the orchestration platform, with the objective of enhancing the experience of the user as well as keeping important metrics in check, including resource utilization and load distribution over the Edge compute nodes. We perform evaluations on a physical testbed to assess the performance of the Edge-based system using real-world video transcoding workloads.
The growing diversity of data sources in modern organizations has made efficient querying and integration crucial for actionable insights. With data distributed across structured relational databases like MySQL and semistructured NoSQL systems such as MongoDB, querying heterogeneous environments has become more complex. Apache Drill and PrestoDB are two prominent query engines designed to address this challenge by providing federated querying across multiple data sources. This paper examines the architecture, design, and performance of Apache Drill and PrestoDB, focusing on their integration with MySQL and MongoDB. Using a synthetic smart city dataset that simulates urban mobility, traffic, air quality sensors, and public infrastructure usage, a series of queries, spanning from basic data retrieval to complex multi-source joins and aggregations, were executed to assess each engine’s performance across various workloads. This research provides insights into the trade-offs between flexibility and performance when querying heterogeneous data environments. It also offers guidance for optimizing data retrieval and analysis across various storage systems.
Breast cancer is the most common cancer in women, with early detection playing a crucial role in improving survival rates. Currently, diagnostic methods largely rely on the manual interpretation of mammograms. The process is time-consuming and prone to human error. This paper explores the development of an automated breast cancer segmentation and classification system using image processing and deep learning techniques. Building on the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) dataset, the system uses a multi-model approach, including a U-Net for tumour segmentation and a Residual Network 50 (ResNet50) based pathology classifier. A MATLAB-based Graphical User Interface (GUI) integration is then presented to radiologists to allow them to interact with the algorithm developed. Uploading mammograms, performing automated analysis, and visualising results is possible. The model evaluation demonstrated strong segmentation performance (99.25% accuracy) and effective classification.
The growing demand for highly reliable, low-latency communication in industrial and real-time applications poses significant challenges for traditional Wi-Fi networks, especially under dynamic traffic conditions. Multi-Connectivity (MC) presents a viable solution to enhance network performance by utilizing multiple transmission paths. This paper extends our previous work where we proposed Channel Capacity Index-Aware Adaptive Scheduler (CIAS), and evaluates static and adaptive scheduling, i.e., selecting which paths are used, on latency and reliability in non-periodic, sporadic, and attenuated scenarios, which are commonly observed in autonomous systems and industrial applications. We analyze the performance of static schedulers and compare them with CIAS, an adaptive cross-layer scheduling approach that dynamically selects the optimal MC scheme based on channel capacity. Our results demonstrate that CIAS consistently reduces latency and mitigates packet delays, particularly under high traffic loads, sporadic transmissions, and signal attenuation. Furthermore, CIAS enhances reliability by minimizing tail-end latencies and adapting to channel fluctuations more effectively than static schedulers. This study underscores the critical role of adaptive MC scheduling in improving wireless communication efficiency for time-sensitive applications that demand both low latency and high reliability.
This work compares univariate and multivariate Long Short-Term Memory (LSTM) models for predicting delays in public transportation. Only historical delay data was used to train the univariate model, which is effective for last-minute predictions but limited in capturing wider temporal correlations because it is optimized for real-time inference with minimal data requirements. To better capture intricate patterns in transportation delays, the multivariate model incorporates extra contextual factors like time of day and topographical coordinates. Multiple multivariate models were trained on different time intervals (e.g., daily, weekly, and monthly) to evaluate the impact of training data selection on predictive accuracy. The findings show that multivariate models provide better long-term accuracy, especially when trained on properly segmented data, whereas univariate models are computationally efficient and excellent for short-term updates. These results suggest how to best adapt LSTM-based forecasting models for dynamic, practical transportation applications.
Delay Tolerant Networks (DTNs) play a vital role in disaster response to address intermittent connectivity, in particular when Unmanned Aerial Vehicles (UAVs) are deployed to relay critical information. However, most existing DTN protocols are highly susceptible to Denial-of-Service (DoS) attacks because their flooding-based or simplistic routing decisions can be exploited by malicious nodes to quickly saturate buffers, exhausting network resources and disrupting legitimate traffic. This paper presents an Epidemic Oracle (EO) implementation, to mitigate DoS threats by intelligently removing delivered messages from all buffers, thus reducing overhead and freeing network resources. Through extensive simulation in the ONE environment, EO is evaluated against three established encounter-based DTN protocols—Epidemic, Spray and Wait, and Spray and Wait Binary—under varying buffer sizes, transmission speeds, and both aggressive and stealthy DoS attacks. The findings indicate that EO substantially increases delivery ratios while curtailing buffer congestion, even under severe adversarial conditions. These improvements highlight the potential of oracle-based interventions to bolster performance in UAV-assisted disaster scenarios, paving the way for more resilient and efficient DTNs in emergency communications.
Vehicular Ad Hoc Networks (VANETs) are increasingly evolving beyond traditional traffic safety applications to support intelligent environmental monitoring across urban and highway settings. As concerns over air pollution, noise pollution, and greenhouse gas emissions rise, utilizing vehicles as mobile sensor platforms offers a promising solution for real-time, largescale environmental data collection. This paper investigates a system in which a vehicle, equipped with an environmental monitoring unit, gathers atmospheric data and transmits it to a cloud server. Three distinct data management strategies are proposed and evaluated, using two communication technolo-gies-GPRS and LTE-to compare their effectiveness. The study focuses particularly on assessing the energy consumption of the overall monitoring system. Experimental results highlight the key differences among the three strategies and the impact of the communication technologies on data transmission efficiency and energy usage.
The widespread adoption of IoT has driven the development of cyber-physical systems (CPS) in industrial environments, leveraging Industrial IoTs (IIoTs) to automate manufacturing processes and enhance productivity. The transition to autonomous systems introduces significant operational costs, particularly in terms of energy consumption. Accurate modeling and prediction of IIoT energy requirements are critical, but traditional physics- and engineering-based approaches often fall short in addressing these challenges comprehensively. In this paper, we propose a novel methodology for benchmarking and analyzing IIoT devices and applications to uncover insights into their power demands, energy consumption, and performance. To demonstrate this methodology, we develop a comprehensive framework and apply it to study an industrial CPS comprising an educational robotic arm, a conveyor belt, a smart camera, and a compute node. By creating micro-benchmarks and an end-to-end application within this framework, we create an extensive performance and power consumption dataset, which we use to train and analyze ML models for predicting energy usage from features of the application and the CPS system. The proposed methodology and framework provide valuable insights into the energy dynamics of industrial CPS, offering practical implications for researchers and practitioners aiming to enhance the efficiency and sustainability of IIoT-driven automation.
Cloud-Edge Computing Continuum (CECC) systems drive digital transformation by linking cloud services to decentralized edge devices, supporting critical applications such as eHealth. However, trust management in such dynamic and distributed federated systems remains a challenge. This paper proposes the Trust Manager, an advanced framework that integrates the IEEE Federation Hosting Service (FHS) model with blockchain technology. It enhances trust through comprehensive trust profiles for infrastructure providers, incorporating performance metrics, Service Level Agreement (SLA) compliance, and user feedback, thus improving scalability, performance, and reliability in federated systems.
