Explainable Artificial Intelligence (XAI) has become an important field of study to enable the exploration of interpretability and transparency issues relating to deep learning models, especially in high-stakes image analytics systems like medical imaging, surveillance, autonomous systems, and industrial inspection. Convolutional Neural Networks ( CNNs ) have been shown to outperform other state-of-the-art models in image classification, image detection, and image segmentation, and Generative Adversarial Networks (GANs) have been pioneers in generating synthetic images and data augmentation. Although successful, the CNN and GAN architectures are frequently criticized as black-box models that restrict the trust of users and regulatory requirements and implementation in sensitive devices. This paper will introduce a single Explainable AI system, which determines explainability mechanisms in CNN-based image analytics and GAN-based image generation. The framework also presented model-level, feature-level, and instance-level interpretability of CNN classifiers through gradient-based attribution, concept activation vectors, and saliency-based analysis of attention. Meanwhile, explainability is inherent in GAN models by discussing latent space representations, generator-discriminator dynamics, as well as the semantic disentanglement of generated elements. The framework permits transparency in predictive decisions, as well as in the generative mechanisms supporting the synthetic creation of images. A modular pipeline is made so that it enables interpretability throughout training, inference, and synthetic data generation phases. Mathematical formulations of CNN feature attenuation and GAN latent variable sensitivity analysis are introduced to give a theoretical basis. Benchmark image datasets are evaluated experimentally to evaluate the accuracy of classification, generative fidelity, metrics of explainability, and human interpretability scores. Findings indicate that the proposed framework is highly effective in enhancing model transparency without affecting the predictive performance or synthetic image quality. This work has made the following contributions: (i) a single explainable architecture with both CNN and GAN models, (ii) formal explainability measures of generative models, as well as (iii) a scalable framework applicable to practical image analytics systems. The study develops credible AI bridging performance and interpretability in contemporary deep learning-based image systems.
Melanoma continues to be one of the most prevalent causes of mortality cases involving skin cancers. Therefore, early diagnosis is vital to help increase patient survival rates. This review article aims to provide in-depth information on deep learning approaches applied to melanoma detection and analyze papers on this subject written between 2016 and 2025. Also, the paper will highlight publicly available databases, novel architectures, and existing difficulties associated with their application. In particular, it will explore the features of three popular datasets (ISIC, HAM10000, and PH2) and analyze their strengths and weaknesses. The focus will be on such factors as the class imbalance problem and the underrepresentation of various demographics. In addition, different architectures used for melanoma detection will be compared, namely VGG, ResNet, Inception, and novel Vision Transformer networks based on their efficiency and capacity to classify melanoma. Finally, major challenges faced during implementation, including overfitting of models, difficulty in their generalization to other populations, and the problem of black box methods, will be considered. At the same time, potential future directions of research, such as the application of explainable artificial intelligence, federated learning techniques, and the use of various demographic groups as training data, will be presented.
It is critical to prioritize tasks in an RTOS for safe and efficient operation of the system. Many engineers utilize informal ad hoc techniques to assign task priorities, resulting in missed deadlines, priority inversions, and potentially catastrophic failures in safety-critical areas. This paper introduces the Criticality Urgency Quadrant Model (QUADRANT), a systematic two-dimensional model for classification and assignment of task priorities with orthogonal axes of urgency and criticality. QUADRANT is intended to give practitioners on standard fixed-priority RTOS platforms without formal mixed criticality analysis infrastructure a lightweight, deployable engineering heuristic. The model has been validated through empirical tasks executed with FreeRTOS (POSIX simulation, Linux 5.15, Intel Xeon) and 16 concurrent tasks performing 300,000 iterations of CPU stress workloads over a 120 second time frame, ultimately producing results that can be compared against Rate Monotonic Scheduling (RMS), Deadline Monotonic Scheduling (DMS), Audsley’s Optimal Priority Assignment (OPA) Method using Response Time Analysis (RTA) and ad hoc. Key results indicated that QUADRANT executed with zero deadline misses across all task quadrants under normal conditions, while ad hoc produced 769 total deadline misses (653 Q1); DMS produced 229 misses; RMS produced 688 total deadline misses; the majority of which were located in the critical Q2 monitor, while OPA/RTA declared the task set un-schedulable due to priority scarcity (16 tasks, 10 priorities). This emphasizes a known characteristic of worst-case analysis: the lack of comparability with the average case for performance. OPA and QUADRANT serve different needs as complementary tools, not as competitive algorithms. Additionally, the evaluation performed through three-level stress testing (baseline 45 seconds, rigorous 90 seconds, heavy 120 seconds) and 122,000 total task executions validates graceful degradation under overload conditions. QUADRANT produced 26 minor Q2 misses (0.065% failure rate) under maximum load; while RMS produced 532 Q2 misses and ad hoc produced 12,470 Q2 misses, evidencing superior assignment and execution of task priorities. The results on the POSIX simulation platform are valid; though actual execution timing may be different when executed on bare-metal targets, the relative effectiveness of priority assignment can be easily replicated between platforms. All source code of the open-source FreeRTOS implementations is provided for reproducible research and adoption within automotive, medical, and aerospace industries.
Every day, millions of people living with visual impairments rely on a simple white cane to move safely through the world. It has served its purpose for decades, but it only tells you what is directly in front of you. Everything else remains invisible: a puddle across the pavement, a sudden drop at a curb, a flight of stairs appearing without warning, or the terrifying moment of an unexpected fall. Rather than offering another incremental upgrade, MIRAI approaches assistive navigation differently. It identifies obstacles as they approach, recognizes water surfaces before contact, detects pits and steps in advance, and registers fall events the instant they happen. Real control is placed back in the user's hands — a deliberate shake gesture triggers an emergency alert immediately, while a single clap helps locate the device if misplaced. When a dangerous situation unfolds, MIRAI does not wait. It automatically sends the user's live GPS location by SMS to a trusted contact, ensuring help can arrive without delay. Clear voice alerts keep every interaction natural and easy to understand. MIRAI is not simply a device. It is a companion built around dignity, independence, and the deeply human need to move through the world without fear.
The growing abundance of big data has made it extremely difficult to find meaningful and accurate information from larger unstructured, semi-structured data collections. Classical extraction methods are limited by being computation-intensive, slow, and not flexible in heterogeneous data sources. This study presents an optimal extraction framework based on generative models such as generative adversarial networks and vibrational autoencoder, which are scalable, maintain accuracy, and processing speed in the presence of a large-scale dataset. Generative learning models are employed to transform unrevealed representations of inputs into a structured and analyzable format, resulting in improved indexing and retrieval accuracy. Python was used as a primary programming language to implement the system. Machine learning and data processing libraries were also used for training the model, preprocessing of data, and evaluating its performance. When compared with existing MapReduce-based methods, results showed that this method enhanced the accuracy of the extraction of data and the effectiveness of search and retrieval. Also, the processing time was reduced in the process.
Lassa fever remains a significant public health concern in West Africa, with an estimated 100,000 to 300,000 cases and over 5,000 deaths annually. Early detection is hindered by diagnostic delays, symptom overlap with other febrile illnesses, and insufficient diagnostic infrastructure. The effectiveness and generalizability of existing Lassa fever prediction models are limited by class imbalance, feature redundancy, and suboptimal configurations, despite the promise of machine learning for disease identification. This study introduces a hybrid framework integrating Edited Nearest Neighbors (ENN) for noise reduction, Synthetic Minority Oversampling Technique (SMOTE) for class balancing, and Ant Lion Optimization (ALO) for feature selection. The framework was evaluated using 20,062 clinical records with 99 features from Nigeria's disease surveillance system collected between 2017 and 2022. When combined with Random Forest classification, the ALO+SMOTE+ENN approach achieved 100% accuracy, precision, recall, F1-score, and an AUC of 1.00. In contrast, conventional methods such as Logistic Regression, Support Vector Machine (SVM), LightGBM, and Gradient Boosting attained only 75–76% accuracy and exhibited notable precision-recall trade-offs, indicating a 24–25% improvement with the proposed method. The superior performance is attributed to Random Forest's robust learning on preprocessed data, ALO's comprehensive feature selection, and SMOTEENN's effective management of the 3:1 class imbalance and noise. This approach reduces diagnostic uncertainty while preserving computational efficiency and clinical interpretability, supporting reliable automated diagnosis in resource-limited healthcare environments. The findings underscore that improving data quality through advanced preprocessing and metaheuristic optimization yields superior results compared to applying complex algorithms to imbalanced datasets, with significant implications for AI-driven infectious disease surveillance in sub-Saharan Africa.
Population-based swarm-intelligence algorithms are metaheuristic algorithms inspired by the behaviour of animals, birds, and insects found in nature. These algorithms are based on a balanced strategy of exploration and exploitation and can be used to optimize multiple NP-hard problems. The Sand Cat Swarm Optimization (SCSO) algorithm is a novel and recently proposed algorithm based on the hunting behaviour of sand cats. These sand cats have extraordinarily low-frequency hearing that helps them locate and catch their prey. The sand cat swarm optimization algorithm has been successfully tested for solving various optimization problems, and it performed efficiently in comparison to other existing algorithms. However, the SCSO has certain limitations when it comes to convergence and optimality of the global solution. The main aim of the paper is to present a detailed description of the implementation of SCSO and highlight its limitations. In addition to this, various approaches proposed to enhance the performance of SCSO are mentioned, and the adopted methodology is analysed. The changes incorporated in the modified versions of SCSO have been tabulated at the end of the paper.
The traditional scheduling techniques used in real-time applications are generally centered around deadline compliance. However, they do not address the critical aspect of data freshness, which significantly impacts the quality and safety of the output from sensor fusion systems where temporal validity is essential. This paper proposes the use of a value-aware scheduling framework that includes an explicit data degradation model and allows for prioritization of tasks based on deadline urgency and temporal value. This framework introduces the following concepts: an exponential degradation model based on the sensors’ temporal validity windows, rejection of jobs that would produce stale data once completed, and dynamic priority assignment based on the sensor’s deadline urgency, time to fresh data, and criticality levels that conform to ISO 26262 safety standards. FDS was successfully implemented using FreeRTOS to evaluate the performance against an automotive sensor fusion application and was subjected to extremely high effective loads (303.6% and 415.6% demand scenarios). Performance results showed between 42–381% performance improvement over traditional fixed priority scheduling, and during the evaluation, no ASIL-D (Automotive Safety Integrity Level D - Highest Level of Safety Required by a Safety-Critical Application) deadlines were missed. Formal analysis provided guarantees of admitted jobs for all ASIL-D tasks, and evidence for an associated reduction in worst-case response times as a consequence of admission control interference. The work challenges the deadline-centric paradigm and shows that timely deadline completion does not guarantee retention of the temporal coherence of the output; rather, value-aware scheduling is required to provide temporal coherence.
In software engineering, software evolution is applied after the initial version of the software has been developed. Also, software evolution is based on the enhancement of the software system by adding new functionalities and correcting defects. Since such updates are complex and prone to errors, systematic methods are required to manage software changes successfully. Software process tailoring means reshaping standard organizational processes to light on project-specific needs while still maintaining compliance and needed verification and validation tasks or activities. Applying traditional process designs and standards supports better quality, decreases risks, and reduces rework in the development of software. In this paper, we consider software evolution processes modeled as basic blocks applied to Petri Nets (PNs) and tailored by four important operations: addition, deletion, splitting, and merging to modify workflows. In tailoring the software evolution process, each basic block is studied using control vectors, incidence matrices, and state equations, with reachability analysis ensuring the required process outcomes. The marked graph PNs represent structured concurrency without uncertainty and conflict, and confluence matrices support structural study by identifying competing and convergent execution paths, thereby protecting correctness and consistency of workflow behavior. Structural analysis shows basic blocks with various executions, i.e., sequence block preserves deterministic control flow, concurrency block represents parallel execution, selection block enforces valid choice paths, and iteration block identifies infeasible loops. According to the diagonal elements of the Conflict and Confluence matrices, one guarantees each activity is consistent and deterministic, allowing safe composition of basic blocks. Together, these results confirm that the Petri Net (PN) based application produces sound, reliable, and feasible software evolution processes. Effectively, PNs' basic blocks model in software evolution, and their matrix representations enable systematic analysis. Hence, Software processes can be tailored by main operations: addition, deletion, splitting, and merging, with consistency between highand low-level models maintained to ensure correctness and reliability.
The high rate of oil and gas digitalization has brought about a lot of complications in the processes of handling occupational health and safety in distributed infrastructures. The conventional safety measures are usually reactive in nature and therefore opening systems to break down in terms of the operations, and also due to human error. That is why it is urgent to preventively combat the risks by combining HFE with the latest digital technologies. It is against this backdrop that this paper proposes a cyber-resilient, cloud-edge platform-based digital twin technology in developing better OHS through realtime monitoring and predictive analytics. The approach is a design-science approach, which includes the integration of a zero-trust approach to access, an adaptive security approach, and an automated self-healing approach to make sure that the System undergoes constant synchronization. A high-fidelity simulated validation shows the high-detection rate of 99.95% of the operational anomalies present in the System and shows that the platform recovered all nodes with a 100% node recovery rate, since it enabled automated failure and recovered system throughput to 85.85%, even after experiencing simulated disruptions. The evidence based on these findings shows that the application of the concept of HFE into the framework of resilient digital systems is highly likely to positively contribute to the process of proactive risk management in high-stakes energy contexts.
The accelerated growth of digital health records, multimodal patient data, and unstructured clinical narratives has overburdened the conventional recommendation systems used in healthcare, and they are incapable of working with complex long-term histories, contextual logic, and multimodal integration. Although Large Language Models (LLMs) have improved natural language understanding and decision support, there are common issues that prevent them, such as hallucinations, insufficient interpretability, safety risks, domain bias, inconsistent reactions, and unreliability across clinical domains, which inhibit the clinical reliability of Large Language Models. This study introduces a hybrid architecture, which is a synergetic integration of LLMs (contextual and reasoning), multimodal modules (clinical image and report analysis), and graph-based collaborative filtering to learn patient longitudinal interactions and collaborative cues. In order to solve hallucinations and uncertainty, the framework involves retrieval-augmented generation, multi-LLM ensemble uncertainty quantification, and knowledge-grounded verification. Tracing paths of reasoning, uncertainty maps, and justifications provided to clinicians to explain explanatory models is built into explainable models to build trust and validation. The system is strictly tested against actual clinical data and known standards (MedQA, MultiMedQA, MEDHALU, and other emerging suites of practice applications such as HealthBench and DiagnosisArena). Early findings show that it is more accurate in diagnostic procedures, has fewer hallucinations (lowering it to less than 2 percent), achieves greater safety in adversarial use, and personalizes better than either standalone LLM or conventional methods. The paper paves the way for creating safe, equitable, and clinically viable decision support tools by filling the knowledge-based reasoning-collaborative longitudinal recommendation gap, which can empower human expertise and not replace it.
Cyber-Physical Systems (CPS) are the foundation of current smart infrastructure, but the growing interoperability of heterogeneous CPS elements makes them highly susceptible to cascading cyber-physical attacks. To overcome this issue, the paper presents a resilience-first security framework that combines data quality improvement, interoperability-conscious dependency modeling, and system-level risk and resilience analysis. The suggested method is tested using the SWaT industrial water treatment dataset, a collection of multivariate time-series data sampled at 1 Hz. CPS subsystems are decomposed and represented in the form of dependency graphs, and anomaly detection is performed through multivariate feature engineering. A CPS-aware data repair mechanism is used to cope with severe data incompleteness. Experimental findings indicate that the proposed repair method reduces approximately 7 × 10⁶ missing values to nearly zero, achieving almost 100% data completeness. Dependency-based risk analysis demonstrates that during an attack, system risk increases to the range of 0.85–0.95 for more than 30,000 samples. Chemical Dosing (≈0.75) and Distribution (≈0.65) are major contributors to risk propagation in subsystem analysis. Resilience evaluation shows rapid detection (1–2 samples), recovery within approximately 25 samples, and a composite resilience index of 0.03, validating the effectiveness of the proposed framework in enhancing CPS robustness and recovery.
This paper proposes the Extended Honey Badger Optimization (EHBO) algorithm, which improves the Honey Badger Algorithm by using an adaptive control parameter to balance exploration and exploitation. EHBO is tested on six benchmark functions and compared with PSO, DE, ABC, BA, and HBA using 30 independent runs. The results indicate that EHBO performs well with smaller variability, especially on multimodal functions. On Rastrigin, EHBO reached a mean fitness of 4.87E+01 (std 8.31E+00), performing better than PSO (5.33E+01, std 9.85E+00) and DE (3.96E+01, std 6.72E+00). On Ackley, EHBO obtained a mean fitness of 1.91E+01 (std 1.84E+00), showing smaller variability than HBA (2.12E-02, std 8.31E-03). Non parametric tests (Wilcoxon, Friedman) at a 95% confidence level verify the statistical significance of EHBO’s improvements. The proposed algorithm ensures stable convergence and less dependence on initial conditions, making it a trustworthy solver for complex optimization problems.
The rapid expansion of IoT ecosystems has increased concerns regarding data privacy, security, and model integrity, particularly in environments vulnerable to sophisticated adversarial and quantum-enabled attacks. Traditional cryptographic and Federated Learning (FL) methods cannot keep up with the demands for confidentiality and robustness in vast, diverse networks. Therefore, this study presents a Quantum-Inspired Cryptographic Framework that uses Quantum Neural Networks, Post-Quantum Cryptography, and secure Federated Learning to enhance intrusion detection systems and provide quantum resilient communication. The methodology uses the UNSW-NB15 dataset, which has been cleaned, encoded, and normalized with reduced features balanced by SMOTE before splitting into 70 percent training and 30 percent testing. A QNN-based FL model can then be trained while Kyber-512 and NTRUEncrypt secure all updates to the model, as well as node communications, against both classical and quantum threats. The experimental results show that the proposed model significantly outperforms classical FL frameworks, achieving an accuracy of 98.1%, precision of 97.5%, recall of 97.9%, and F1-score of 97.7%, even under very challenging adversarial conditions. QICF is therefore a robust, privacy-preserving, and attack-resilient solution for next-generation large-scale IoT networks.
DevOps teams face an ever-growing challenge in managing log and metrics data: how long to retain data to balance operational value against storage costs and performance constraints. Traditional static retention policies struggle to cope with explosive data growth and evolving compliance requirements. In this work, we propose an intelligent data retention recommendation system that leverages Elasticsearch’s rich monitoring data and Machine Learning (ML) to suggest optimal retention periods for indices dynamically. Our approach collects metrics on query load, storage use, and index lifecycle policies from a live Elasticsearch cluster and trains an ML model to predict the retention duration that minimizes cost while preserving necessary data availability. We present a framework where the model learns usage patterns and system constraints, recommending when to tier or delete indices. Preliminary evaluations suggest that the ML-driven approach can reduce storage costs and cluster strain by avoiding over-retention of seldom-accessed data, without compromising on query performance or compliance. This paper details the related work in intelligent log management, the theoretical underpinnings of our approach, the design of our ML-based retention recommender, and experimental results in a DevOps context. We conclude with insights into the benefits of adaptive data retention and discuss future improvements for integrating such systems into automated DevOps pipelines.
This article introduces a new scientific discovery in the field of computational number theory: an enhanced integer factorization method that unifies an optimized form of a prime generator algorithm with an advanced trial division framework. This integration produces a factorization technique that is both faster and more structurally elegant than conventional approaches. The discovery lies in demonstrating how a modern prime generator can be adapted and expanded to drive the factorization process with greater precision, reducing unnecessary computations and improving overall performance. The paper details the conceptual foundations of the method, outlines its computational advantages, and explains the process choices that make the algorithm both practical and theoretically elegant. The article describes the transformation of a simple prime generator algorithm into a solution for an NP problem, achieving optimized time and space efficiency.
Conventional object detection models are fully supervised and rely on large-scale labeled datasets with bounding box annotations for each object category. However, collecting such labelled datasets for every possible class is costly and impractical. To address this drawback, the proposed method uses a novel Zero-Shot Object Detection (ZSD) framework that detects unseen object categories using only natural language descriptions, without requiring additional training or labelled data. The method integrates CLIP, a vision-language model trained on image-text pairs, with a prompt-aware Region Proposal Network (RPN). The RPN is conditioned on CLIP’s text embeddings, enabling it to generate proposals that are semantically aligned with the given text prompt. During inference, the Model compares text and image features in a shared embedding space, allowing it to localize and classify previously unseen objects. Experimental results on the COCO and LVIS datasets demonstrate that our approach achieves competitive performance under zero-shot settings, effectively generalizing to novel object classes based solely on textual input.
Machine learning is significant in clinical decision support, especially regarding dermatological disorders, whose symptoms significantly overlap, and whose features are mostly discrete. The Logistic Regression (LR) can be employed in such a case, as it gives clear probabilistic results with which a clinician can work, but the literature in this field rarely presents a systematic comparison between LR and other classical algorithms in consistent assessment environments. This paper analyses the LR with the Dermatology dataset on the WEKA system and compares the performance with Naive Bayes, Support Vector (SMO) machine, and Decision Tree (J48). This work is novel due to its interpretability-based evaluation and the systematic analysis of classical models with the help of a discrete and symptom-based dataset of dermatology, a field where journal publications and literature tend to prioritize the precision of data over transparency or rigorous methods in data analysis. The study makes use of 10-fold cross-validation to examine the predictive accuracy and error behavior of each model in order to gain insight into the diagnostic potential of each model. Results indicate that LR has competitive accuracy and stable performance, but is interpretable, which is crucial in clinical applications. Even though SMO has a minor difference in accuracy, LR offers reliable and interpretable predictions, which apply to structured clinical data. The results provide an essential methodological foundation and emphasize the applicability of interpretable models to dermatology diagnostics, where there is a clear, understandable line of decision-making that can aid credible and reliable automated services.
