Zero-knowledge proofs (ZKPs) enable a prover to convince a verifier of knowledge of a secret without revealing it. The ZKP for the square-root problem has many applications in network and cloud security, such as user authentication and privacy-preserving cloud storage auditing. Classical protocols for the quadratic residuosity (square-root) relation require multiple iterations to reach negligible soundness error, incurring latency and communication costs that are critical in cloud settings. This paper proposes a new single-round zero-knowledge proof (SR-ZKP) for the square-root problem that achieves the same soundness as iterative schemes by increasing the challenge length. The protocol requires only one execution of a 4-message protocol (request, commit, challenge, response) and can be transformed into a one-message non-interactive ZKP via the Fiat-Shamir heuristic. The completeness, soundness, and zero-knowledge properties of the proposed scheme are formally proven. The results of this study show that the proposed protocol can achieve approximately 97% reduction in communication overhead and latency, when compared to an 80-round iterative ZKPs with RSA modulus.. of size 2048 bits. This provides a substantial advantage for cloud applications.
Evasive poisoning attacks such as semantic backdoors pose a growing threat to federated learning because they mimic benign client updates and evade detectors under secure aggregation. We introduce an unsupervised, per-client structural check that runs after each local round, before aggregation, requiring only compact statistical summaries derived from each client update after a geometric transformation that removes dependence on the global model. Client-update statistics are compared against a calibrated benign reference, and deviations are detected through statistical distances. Energy and Wasserstein-1 jointly define an operational pattern where threshold exceedances across projections reveal structural deviations even in apparently benign updates. Evaluated on canonical backdoor scenarios from Bagdasaryan et al., the method detects both strong and stealthy attacks in the first local round through consistent multi-projection threshold excesses, while benign updates show only isolated ones. The procedure is lightweight, unsupervised, compatible with secure aggregation, and does not require trigger datasets. By providing early per-client warnings before aggregation, it complements classical defenses such as norm clipping, differential privacy, and robust aggregation, enabling proactive mitigation of poisoning in federated learning.
This study constructed a dataset targeting character product images, that are an important source of information about customer behavior, for business marketing. The images were collected independently and assigned detailed tagging information based on hypotheses, with the aim of training the models and contributing to multifaceted evaluation. Additionally, experiments using classification models demonstrated that the constructed dataset and tagging information contributed to verifying the characteristics of the models.
Encrypted traffic classification has emerged as a critical component of modern network management and cloud security services. While Virtual Private Networks (VPNs) ensure user privacy by encrypting communications, this encryption also complicates traditional traffic identification. Recent research demonstrates that Machine Learning (ML) and Deep Learning (DL) techniques can effectively classify VPN versus non-VPN traffic even without payload inspection. However, to align with trustworthy cloud service requirements, these ML/DL approaches must also preserve user privacy and assure security. In this paper, we review the state-of-the-art ML/DL methods for encrypted VPN traffic classification, emphasizing techniques that enhance trust, including privacy-preserving federated learning, adversarially robust models, and explainable AI. In addition, this work aims to discover the most significant features affecting the VPN classification and identifying the best-performing ML and DL models on available VPN classification datasets. We include studies that focused on characterizing the VPN traffic besides classifying the secure traffic into VPN and non-VPN.
This work presents a dual-modality unmanned aerial vehicle (UAV) detection and classification approach designed to address challenges in aerial surveillance under low-visibility and cluttered environmental conditions. The core contributions of this work are twofold. First, we introduce a novel, high-quality multimodal dataset comprising both real-world and synthetically generated UAV imagery across red-green-blue (RGB) and infrared (IR) channels. The dataset was curated through a rigorous pipeline that includes multi-source video collection, frame extraction, noise reduction, simulation-based class balancing, and scene-aware clustering to prevent data leakage. Second, we present a comprehensive benchmarking of detection and classification models using this dataset. Detection models were finetuned using You Only Look Once Version 11 Nano (YOLOv11n), one for RGB and another for IR, achieving mAP@0.5 scores of 93.5% and 98.2%, respectively. The proposed approach also performs classification into fixed-wing and rotary-wing UAV types to support more informed threat evaluation. Integration of the Slicing Aided Hyper Inference (SAHI) framework increased true positives by 24.7% (RGB) and 2.4% (IR), significantly enhancing detection performance.
In this paper, we propose an approach to solve an "use-inspired" problem: the management of warehouses containing equipment to be installed at various sites throughout the territory by specialized personnel. The challenges to be addressed include the distribution of the warehouse, so as not to slow down work, but also the fact that the tasks to complete change constantly, requiring the opening, closing, and relocation of the corresponding warehouses. The study models the problem as an optimization problem and investigates how to compose multiple objectives to select the warehouse location, in a single objective function, necessary to obtain a single solution, and evaluates the performance of a set of stochastic optimization algorithms to generate cost-effective solutions while adhering to the constraints of the problem. The results demonstrate that metaheuristic algorithms are highly effective in identifying optimal solutions, with Binary Particle Swarm Optimization and Population Based Incremental Learning having good performances.
Reliable water quality monitoring is critical for safeguarding public health and ensuring the sustainability of ecosystems, especially in regions facing growing environmental and industrial pressures. This paper presents a novel edge-intelligent framework that combines Graph Neural Networks with real-time IoT sensor deployments to predict and monitor multiple water quality parameters. Leveraging a two-stage spatio-temporal graph construction process grounded in Euclidean and correlation-based criteria, we model the spatial relationships and temporal dynamics of diverse water parameters. Our models achieve strong predictive performance across 14 critical parameters, including temperature, dissolved oxygen, conductivity, and microbial indicators, with R2 scores as high as 0.92. Deployed on low-cost Raspberry Pi-based edge devices, our system enables real-time inference and energy-efficient operations without reliance on cloud connectivity. This work bridges the gap between deep learning and in-situ environmental monitoring, demonstrating an IoT approach for data-driven water governance. The proposed solution holds significant promise for policy-makers, researchers, and communities aiming to decentralize environmental sensing and address the global challenge of clean water access.
Knowledge Graph (KG) processing faces critical infrastructure challenges in selecting optimal NoSQL database paradigms, as traditional performance evaluations rely on static benchmarks that fail to capture the complexity of real-world KG workloads. Although the big data field offers numerous comparative studies, in the KG context DBMS selection remains predominantly ad-hoc, leaving practitioners without systematic guidance for matching storage technologies to specific KG characteristics and query requirements. This paper presents a KG-specific benchmarking framework that employs connectivity density, scale, and introduces a graph-centric metric, namely Semantic Richness (SR), within a four-tier query methodology to reveal performance crossover points across Document-Oriented, Graph, and Multi-Model DBMSs. We conduct an empirical evaluation on the FAERS adverse event KG at three scales, comparing paradigms from simple filtering to deep traversal, and provide metric-driven, evidence-based guidelines for aligning NoSQL paradigm selection with graph size, connectivity, and semantic richness.
Although numerous hardware and software solutions exist for energy monitoring, they often require extensive setup, configuration, and technical expertise. This paper presents a preliminary study investigating the potential of widely available software traces to estimate software energy consumption. Experiments across synthetic and realistic workloads show that certain hardware metrics and structured software logs, such as SQL statements and Docker container logs, consistently correlate with energy usage, whereas unstructured logs, including general HTTP or Maven logs, exhibit less reliable correlations. These findings suggest that combining hardware metrics with structured software traces, while accounting for workload characteristics and application structure, offers a promising approach for estimating software energy consumption without dedicated energy monitoring tools.
In this work, we investigate the influence of image resolution on contour detection pipelines of surface damage on glass plates in a numerical experiment. For this, the MOSES-Impacts algorithm, which was specially developed for the detection of damage caused by laser-accelerated fragments in laser-matter experiments, is used. The aim is to quantify the performance of contour detection of the algorithm as a dependency of image resolution and to investigate whether low-resolution images are suitable for the determination of repair parameters. Low-resolution images offer advantages: They require less memory and can be sampled and analysed more quickly. The results show that the deviation of the contour shape and area using the MOSES-Impacts algorithm increases with decreasing resolution along a power function. Furthermore, a resolution-related relative deviation of 5% in the detection of the recycling radius occurs at 6419 ppi. Based on the absolute detection errors determined, two strategies for sustainable respectively effective recycling are developed, which could be used in high-power laser systems in the future.
Phishing attacks remain a persistent threat to the confidentiality and trust of cloud environments, enabling credential theft and unauthorized access to sensitive resources. This paper presents PhishDefender, a multi-layer phishing defense framework that enhances trustworthy cloud services through the integration of ensemble machine learning, policy enforcement, and threat intelligence validation. Built on the UCI Phishing Website dataset, the ensemble model combining Logistic Regression, Random Forest, Gradient Boosting, AdaBoost, XGBoost, Multilayer Perceptron and Deep Neural Network achieved 97.82% accuracy, 97.91% precision, 97.74% recall, 97.82% F1-score and a ROC-AUC of 0.988, with an average inference time of similar to 1.05 seconds. These results demonstrate high separability between legitimate and phishing URLs while maintaining practical performance for deployment in real-time cloud applications. The framework further extends detection outcomes into actionable policy responses (Allow, Alert, Report, Block) verified against external threat feeds, forming a layered defense aligned with zero-trust architecture principles. Its lightweight and modular design enables deployment on standard or cloud-hosted infrastructure, offering a reproducible and scalable approach for organizations seeking to enhance trust, resilience, and compliance in distributed cloud ecosystems.
Large Language Models (LLMs) inference-time privacy focuses mainly on protecting sensitive information within the prompt. However, enabling end users to repudiate the submission of exact prompts, to avoid potential legal consequences, remains an unexplored research direction. To address this gap, we propose Repudiation for Good (R4G), a privacy-preserving approach that enables users to deny submitting exact prompts while allowing cloud-based LLM providers to receive sufficient semantic information but preventing them from proving with certainty which specific prompt was submitted. R4G is a paradigm shift in LLM inferencetime privacy. Rather than encrypting exact prompts, which eventually reveal their precise content upon decryption, R4G transforms prompts into an intentionally ambiguous semantic representation through LLM embedding. This transformation creates a many-toone mapping where multiple lexically distinct prompts collapse into a similar embedding space. The key insight is that while traditional privacy approaches aim to hide information temporarily (through encryption) or completely (through anonymization), our approach provides a probabilistic association between the prompt and its embedding, creating a deniability space, where users can legitimately deny submitting exact prompts that might carry legal, social, or professional consequences. We conducted experiments to assess the efficacy of R4G by measuring the semantic similarity between the original and approximated prompts. The results show that R4G achieved an average cosine similarity score of approximately 50%, effectively striking a balance between utility and privacy.
Accurate future workload prediction is an essential step for proactive resource allocation and efficient provisioning in cloud computing environments. Deep learning strategies have proven successful for this task, but they face challenges due to the high dimensionality of monitoring data, extensive preprocessing requirements, and computational overhead. In this paper, we propose a hybrid framework that integrates autoencoders for workload compression with Long Short-Term Memory (LSTM) networks for time-series forecasting. Unlike prior studies, our approach systematically analyzes the trade-off between compression ratio and predictive accuracy, demonstrating how dimensionality reduction can improve both scalability and robustness. Thereby reducing the computational burden associated with processing massive-scale monitoring data. Experiments conducted on both synthetic and real-world datasets demonstrate that the proposed method achieves up to 60% data compression with minimal reconstruction loss, while also improving prediction accuracy compared to baseline LSTM models. We evaluate the overall performance of the framework using various metrics, including data reduction ratio, prediction accuracy, and the effects of different compression stages on predictive performance. Additionally, we quantify the computational savings in terms of CPU usage, memory footprint, and training/inference times, confirming the framework's feasibility for real-world deployment. These results underscore the potential of integrating compression and prediction to achieve scalable, accurate, and resource-efficient management of cloud workloads.
Phishing attacks represent one of the most serious crimes of the 21st century posing significant challenges for researchers in both academia and industry. These attacks are to steal sensitive information from Internet users, such as bank accounts or credit card credentials. Therefore, phishing costs Internet users millions of dollars every year. However, improving user security awareness represents an effective way to prevent these attacks, especially given the significant growth of online services. Therefore, this paper proposes a novel approach to implement an Analyzer Model for phishing detection. This approach is set up to automatically and continuously analyze user responses to phishing attacks and enforce training interventions. This approach was evaluated in a real-world experiment targeting 207 users in the education field. The results of this study reveal that phishing awareness has a significant positive impact on users' ability to distinguish between phishing emails and websites, thereby avoiding such attacks.
Reliable detection of embryo developmental transitions is central to clinical decision-making in in vitro fertilization (IVF), yet current practice still relies on subjective visual inspection by embryologists. This study investigates how to scale from proof-of-concept AI models to deployment-ready systems by benchmarking eight spatio-temporal deep learning backbones for transition boundary detection in time-lapse embryo imaging. Using a leakage-free embryo-level split on a public IVF cohort, we model short-term temporal context through 8-frame temporal stacks from the central focal plane-referred to as a 2.5D representation and evaluate a unified training and inference pipeline across EfficientNet-B4, DenseNet121, VGG16, Inception-v3, ViT-B/16, Swin-T, AlexNet, and ConvNeXt-Tiny. We pair predictive metrics with system-level indicators: number of parameters, FLOPs, model size, and inference latency, to capture cloud-edge deployment constraints. Empirically, Swin-T achieves the best overall macro performance (Accuracy = 69.91%, F1 = 0.6989), followed closely by ViT-B/16, while EfficientNet-B4 provides the most balanced accuracy-efficiency trade-off (F1 = 0.6663 at 1.53 GFLOPs). Lightweight CNNs such as AlexNet and ConvNeXt-Tiny minimize latency (sub-ms to fewms) but underperform in predictive accuracy. Pareto analyses (F1 vs FLOPs/latency/parameters) reveal a consistent efficient frontier that supports a tiered deployment strategy: lightweight CNNs for real-time edge screening, Transformers for cloud-scale batch analytics, and EfficientNet-B4 as a robust hybrid baseline. Overall, the proposed benchmark yields a relative improvement of +12% in macro F1 over prior baselines and establishes a reproducible reference for evaluating accuracy-efficiency trade-offs in IVF morphokinetic analytics. The released benchmark, ETL pipeline, and ablation guidelines provide actionable insights for scaling embryo transition detection across the cloud-edge continuum.
Collaborative Learning (CL) grows in relevance within Industry 4.0, as it facilitates knowledge exchange among networked participants and integrates learnings from models into local manufacturing processes. This allows organisations to efficiently harness information from distributed data, reduce costs, and accelerate innovation. Existing research focuses on conventional CL strategies, wherein model aggregation is performed on a central server. This centralised architecture, however, introduces inherent vulnerabilities, including a single point of failure and privacy risks. Recent advancements indicate a paradigm shift toward decentralised collaborative learning (DeCL) systems, yet no unified set of requirements has been proposed for supporting their development and implementation. This study addresses the aforementioned research gap by conducting a systematic literature review. Drawing on a synthesis and analysis of existing studies, this review identifies a comprehensive set of 23 system requirements, organised into six key categories, specifically focused on designing DeCL systems in manufacturing. Future research should align with Industry 5.0 principles by incorporating the human-centric design for ethical artificial intelligence and worker inclusion, as well as incorporating sustainable, resource-efficient technologies to develop environmentally responsible DeCL systems.
Anomaly detection based on time-series environmental data plays a critical role in ensuring accurate and reliable Digital Twin (DT) operations. This study investigates the effectiveness of motif-based approaches to support anomaly detection for humidity data, in order to identify the most suitable solutions for DT-enabled monitoring in smart buildings. We compared eight motif-based methods Matrix Profile, Template Matching, Seasonal Motif, Sliding Window, Multi-Scale, Clustering, Ensemble, and Peak-based - using 30-day intervals of indoor humidity data. We evaluated the performance of each method using a standardized preprocessing pipeline with uniform intervals, interpolation, normalization, and daily segmentation. Based on these, we reported the number of anomalies, detection rates, variability, confidence scores, and temporal patterns. The results show that no single method is universally superior: Seasonal Motif and Peak-Based excel at recurring low-humidity anomalies, Multi-Scale and Clustering capture broader deviations, and Ensemble methods maximize precision but miss subtler events. These findings offer practical guidance for anomaly detection in Digital Twin settings and highlight the importance of expert-labeled validation and hybrid approaches.
Cloud computing offers a scalable infrastructure that gives the impression of virtually unlimited computing resources, attracting a wide range of applications. However, it raises significant concerns regarding data privacy and risks user trust due to making the required computation on a remote machine (rather than on a local machine) and transmitting data over communication channels. Homomorphic encryption provides feasible solutions but introduces extra computational costs due to related cryptographic data conversions. So, it is crucial to balance computational efficiency with effective privacy preservation. We propose an intelligent framework capable of classifying text data into private and public categories, thereby enabling context-aware routing of data processing tasks. Specifically, private data is processed locally to safeguard user privacy, while public data is offloaded to powerful cloud servers for resource-intensive analytics. This selective processing strategy using an efficient machine learning strategy seeks to balance privacy, latency, and efficient resource utilization. Our developed framework was tested with a cloud application using Transformer-based models, specifically DistilBERT that shows strong performance, achieving 99.98% accuracy in private-public classification. Finally, DistilBERT was used for a privacy-aware spam detection cloud application with successful integration of federated learning (FL) approach.
The ongoing surge of ransomware has underscored the need to shift from perimeter-based security to Zero Trust models. This paper investigates a Zero Trust Architecture (ZTA) approach to containing ransomware in a virtualized environment using least-privilege controls, micro-segmentation, and continuous monitoring. We develop an open-source, lightweight security architecture comprising Wazuh for real-time auditing and alerts, audited for system logging, and the Uncomplicated Firewall (UFW) for network segmentation within a VirtualBox laboratory network, consisting of Ubuntu as the victim and Kali as the attacker virtual machines. A simulated ransomware attack is conducted to evaluate detection latency, data impact, system overhead, and alert accuracy. The prototype ZTA framework detected ransomware activity in an average of approximate to 5.3 seconds. This detection limited encryption to approximately 20% of files prior to the activation of containment measures, while maintaining minimal CPU and memory overhead and exhibiting a low rate of false positives. These findings illustrate the successful early containment of ransomware via the implementation of Zero Trust controls. Although evaluated in a laboratory environment, the methodology is applicable to trustworthy and secure cloud or hybrid systems by improving data protection, facilitating compliance-oriented audits, and minimizing the impact of attacks.
The rapid adoption of AI systems capable of autonomous code generation and infrastructure orchestration introduces new challenges for authorization. Traditional mechanisms such as Role-Based Access Control (RBAC), Attribute-Based Access Control (ABAC), OAuth 2.0 and OpenID Connect were designed for static or user-initiated interactions and thus fail to ensure both least privilege and operational continuity in self-adaptive, agentic environments. To address this gap, we propose an agent-native authorization model in which specialized IAM agents-Analyze, Impact, Minimize, Validate, and Reanalyze-jointly determine, verify, and reduce permissions through a sequential probabilistic process. The framework combines Bayesian belief updates, MDP optimization, and CVaR minimization to achieve least-privilege under uncertainty. Formal analysis and an AWS Lambda maintenance case study show sub-second decision latency and similar to 40% privilege reduction without loss of functionality.
