Modern Artificial Intelligence (AI), and Machine Learning (ML) models are widely used in numerous domains. However, their decision-making-based processes are complex for human understanding and interpretation. Such lack of transparency creates serious challenges, especially in sensitive domains as trust is essential. This raises the importance of Explainable Artificial Intelligence (XAI) methods to address such issue by providing tools that explain how AI models reach their decisions. One of the most dominant XAI methods is the Local Interpretable Model-agnostic Explanations (LIME). It explains individual predictions using simple local models. Nevertheless, it exhibits limited fidelity, global interpretability, and varying behaviour according to the used data types. Various studies have proposed improvements to LIME in different applications, which raise the difficulty to develop advanced, generic LIME method. This inspired this work to present a comprehensive survey of LIME-based explanation methods with structured data. It summarizes LIME’s limitations, reviews its modified versions, and provide a unified overview of the progress of LIME. Compared with previously published surveys on XAI and LIME, the proposed taxonomy systematically organizes traditional and modified LIME approaches according to the core limitations toward providing an integrated LIME framework that simultaneously improve stability, fidelity, and global interpretability. The survey also identifies key research gaps, and discusses future directions for LIME-based explanation approaches. The recommended future research took benefit from developing explanation frameworks that address the major limitations of traditional LIME within a unique framework.
Accurate detection of micro-lesions in Multiple Sclerosis (MS) MRI remains a persistent clinical challenge due to extreme class imbalance, low boundary contrast, and the computational demands of existing volumetric deep learning frameworks that limit their deployment in resource-constrained environments. This study proposes a lightweight multi-modal attention U-Net specifically optimized for micro-lesion segmentation, targeting lesion instances with volumetric extent smaller than 10 mm³. The proposed architecture employs a 2D encoder-decoder backbone augmented with micro-lesion-scale attention gates that selectively fuse FLAIR and T1-weighted feature representations, restricting the effective receptive field to sub-centimeter regions to suppress background activations and enhance small lesion sensitivity. Training is guided by a composite loss combining Dice loss, micro-lesion class weighting, and a boundary-sensitive distance transform term, supported by an enriched batch sampling strategy that ensures adequate micro-lesion representation throughout optimization. Evaluation on three publicly available benchmarks — ISBI 2015, MICCAI 2016, and 3D-MR-MS — demonstrates that the proposed model achieves a Dice Similarity Coefficient of 74.6% on the micro-lesion subset, representing a 10.8% improvement over the standard U-Net baseline, while operating at 6.3× faster inference speed and requiring 8.7× fewer parameters than Trans-U-Net++. These results confirm that task-specific lightweight design yields a clinically meaningful accuracy-efficiency trade-off, rendering the proposed framework suitable for real-time deployment in routine MS assessment pipelines.
Single-image super-resolution (SISR) aims to recover a high-resolution (HR) image from a single low-resolution (LR) input, a task with critical applications in medical imaging, satellite surveillance, digital archiving, and multimedia restoration. While generative adversarial network (GAN)-based methods have substantially advanced perceptual quality in SISR, existing models such as ESRGAN still suffer from training instability, sensitivity to loss weighting, and limited robustness to real-world degradations. This paper proposes an enhanced ESRGAN framework that addresses these limitations through three complementary contributions: (1) an extended generator architecture incorporating five additional Residual-in-Residual Dense Blocks (RRDBs) beyond the original 23; (2) a six-component multi-loss formulation combining pixel, VGG-based perceptual, LPIPS, frequency-domain (FFT), edge-aware, and adversarial losses; and (3) a dynamic three-stage training schedule that linearly interpolates loss weights from a content-only warm-up (iterations 1–50k) through a stabilization phase (50k–150k) to full multi-loss training (150k–200k). Evaluated on Set5, Set14, BSD100, and DIV2K at ×4 upscaling, the proposed model achieves a statistically significant mean PSNR gain of 0.775 ± 0.126 dB over ESRGAN (t(3) = 12.32, p < 0.01) and a 16.1 % relative improvement in LPIPS (0.155 → 0.130), confirming both numerical accuracy and perceptual quality gains. Ablation experiments confirm the independent contribution of each added component. The framework retains single-pass inference, offering a practical efficiency advantage over diffusion-based methods. Keywords: Super-resolution; ESRGAN; image restoration; generative adversarial network; perceptual quality; multi-loss design; Residual-in-Residual Dense Blocks; dynamic training strategy.
Waste recycling is essential for a healthy environment and for building a sustainable economy. This paper proposes a hybrid deep learning model that combines deep transfer learning with eXtreme Gradient Boosting (XGBoost) to classify trash in the TrashNet dataset, which contains six categories of garbage images. The proposed approach aims to improve classification accuracy and streamline sorting for sustainable waste management. The model’s performance is evaluated both quantitatively and qualitatively against state-of-the-art techniques. In this study, fine-tuned VGG16, Inception-v3, and MobileNet are integrated with XGBoost classifiers to enhance feature learning and improve classification results. The proposed hybrid models show strong performance, achieving accuracies of 89.72%, 92.08%, and 93.86%, respectively. In addition, the model’s reliability is validated through Grad-CAM and Grad-CAM++ saliency maps, which provide visual evidence of the regions used for classification. These results demonstrate the effectiveness of the proposed method and confirm its potential as a robust and automated solution for improving waste classification and supporting efficient waste management.
Federated Learning (FL) enables collaborative model training while preserving data privacy, but its decentralized nature exposes it to backdoor attacks, where malicious clients inject poisoned updates that embed hidden triggers into the global model. This threat is significantly amplified under non-IID data distributions because natural client heterogeneity can conceal malicious updates. This survey provides a structured analysis of backdoor defense mechanisms in FL under non-IID conditions. A unified taxonomy is introduced, categorizing existing approaches into robust aggregation, trust-based filtering, differential privacy, geometric analysis, and hybrid multi-layered frameworks. Unlike prior surveys that focus narrowly on attack methodologies or assume IID data, this work presents the first dedicated analysis of backdoor attacks specifically under non-IID heterogeneity. Three state-of-the-art hybrid architectures, GRASP FL, SHIELD FL, and HyGrad, are examined, highlighting their effectiveness against backdoors under label skew, quantity imbalance, and feature heterogeneity. A comprehensive evaluation framework covering security (attack success rate, false positive and false negative rates), utility, and efficiency is proposed. Key backdoor-specific challenges are discussed, including adaptive adversaries, the privacy-security-utility trilemma, scalability, fairness, and evaluation standardization. Finally, future research directions are outlined, emphasizing large-scale validation, model heterogeneity, and formal certification. This survey serves as a reference for researchers and practitioners seeking to understand and mitigate backdoor threats in practical non-IID federated learning deployments
Link prediction in dynamic contact networks is a vital task in modeling and controlling the spread of infectious diseases. However, semantic redundancy—where nodes with similar meanings appear under different surface forms—often distorts the underlying graph structure, reducing the effectiveness of graph-based learning. This paper introduces a Semantic Repetition Sampling (SRS) module designed to mitigate this issue by leveraging cosine similarity to identify redundant node pairs. A sampling correction matrix is then constructed to adjust the adjacency structure, enhancing the relevance of training samples. The proposed framework integrates this SRS module into a two-layer Graph Convolutional Network (GCN), forming a lightweight yet effective architecture for epidemic link prediction. Experiments conducted on a COVID-19 dataset demonstrate the superiority of the GCN + SRS model over both the GCN baseline and DLPCov methods. The model achieves an AUC of 0.9287 and a MAP of 0.0918, outperforming the GCN baseline (AUC: 0.9128, MAP: 0.0827) and providing competitive results compared to DLPCov. These findings highlight the importance of semantic filtering in improving graph-based epidemic forecasting and offer a practical contribution to public health analytics.
The integration of Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has significantly enhanced the capabilities of Digital Twins (DTs) by enabling predictive modeling and adaptive intelligence. Collectively, these technologies transform DTs into intelligent platforms capable of adaptive control, predictive analytics, and automated decision-making. This systematic review synthesizes findings from recent literature to examine the extent to which ML and DL are applied across multiple scales, ranging from component-level diagnostics to enterprise-level operational management. Applications span manufacturing, healthcare, civil infrastructure, energy systems, and smart cities, where AI-enhanced DTs support predictive maintenance, patient monitoring, energy optimization, and urban planning. The review analyzes common ML/DL architectures and integration patterns, including simulation-to-real transfer, physics-informed neural networks, and online reinforcement learning. Furthermore, it discusses emerging research directions such as self-evolving cognitive twins, generative AI integration, and federated learning, while critically examining current limitations that constrain the full realization of AI-enhanced Digital Twin systems. This work provides a comprehensive reference for researchers and practitioners seeking to advance the design and deployment of intelligent Digital Twin architectures.
Federated Learning (FL) enables distributed model training without exposing raw data; however, it remains highly vulnerable to backdoor attacks, particularly under non-IID data distributions where statistical heterogeneity complicates reliable detection. This work introduces HyGrad, a hybrid gradient-adversarial defense framework that integrates four synergistic components: (1) gradient affinity filtering based on cosine similarity and density-based clustering, (2) adversarial trust scoring derived from prediction stability under perturbations, (3) temporal trust adaptation via exponential decay, and (4) bias-resilient aggregation using KL-divergence weighting. Operating without access to local data, HyGrad dynamically quantifies client trust, effectively isolating malicious updates while preserving benign contributions, even against adaptive adversarial strategies. Experimental evaluation on CIFAR-10, MNIST, and GTSRB demonstrates that HyGrad maintains high classification accuracy (86.2%, 96.7%, and 82.1%, respectively) while reducing attack success rates by over 90% compared to unprotected baselines, significantly outperforming existing defenses such as CRFL and Bulyan. The framework exhibits practical efficiency for networks of up to 100 clients, requiring approximately 85 seconds per round on CIFAR-10, with scalable extensions enabled through mini-batch clustering approximations. Furthermore, theoretical analysis establishes convergence guarantees under adversarial conditions. These results highlight HyGrad as a robust and practical defense mechanism for secure and scalable federated learning in adversarial environments.
In machine learning, dealing with imbalanced datasets remains a significant challenge. Class imbalance arises when the distribution of instances across classes is uneven, which can occur in both binary and multiclass problems with varying imbalance ratios. Conventional strategies like resampling and reweighting have been commonly used to address this issue but often yield suboptimal performance and limited effectiveness. Many standard classifiers tend to bias their predictions toward the majority class, leading to reduced accuracy, particularly in recognizing minority class instances. Ensemble learning methods have shown promise in improving classification outcomes on imbalanced data. Nonetheless, these approaches often introduce considerable computational costs, extended training times, and scalability issues due to the large number of models involved, which can burden system resources. To address these limitations, this research proposes the use of pruning techniques within ensemble classifiers. By retaining only the most effective classifiers, the ensemble’s size is reduced without compromising— and potentially improving— its predictive accuracy. Experimental results indicate that the proposed method offers a viable alternative to large ensembles, producing smaller, faster, and more accurate models. The resulting pruned ensembles achieve competitive performance, with accuracy reaching up to 99%, while significantly lowering computational overhead, making them well-suited for large-scale applications.
Automated Program Repair (APR) represents a transformative approach in software engineering that aims to reduce manual debugging efforts by automatically generating patches for software bugs. The recent emergence of Large Language Models (LLMs) has revolutionized this field, offering new capabilities for understanding and generating code repairs with minimal human intervention. This paper provides a comprehensive analysis of the current state of LLM-based APR techniques, examining their effectiveness, challenges, and future directions. Based on a review of recent publications and empirical studies, we identify four primary paradigms in LLM-based APR: fine-tuning, prompting, procedural pipelines, and agentic frameworks. Our analysis reveals that specialized code models often outperform larger general-purpose models, with repair performance not scaling linearly with model size. The study further demonstrates that LLM-based approaches can repair 27.78% more bugs on benchmarks like Defects4J compared to traditional APR methods, while successfully addressing complex, multi-hunk, and cross-file bugs that previously resisted automation. However, critical challenges remain in verification beyond test suites, repository-scale defect repair, and computational efficiency. This research synthesizes current achievements, identifies persistent gaps, and outlines promising avenues for future work in this rapidly evolving domain.
Precise segmentation of liver and tumor regions in computed tomography (CT) images is crucial for accurate diagnosis, treatment planning, and surgical guidance. This study proposes a Hybrid DenseNet-U-Net model specifically designed for liver and tumor segmentation. The architecture integrates DenseNet121 as the encoder backbone to exploit dense connectivity for enhanced feature reuse and efficient gradient propagation, while the U-Net decoder performs systematic upsampling with skip connections to preserve spatial details. Additional components such as batch normalization, dropout layers, and optimized filter configurations are incorporated to improve segmentation accuracy and robustness. The model was evaluated on the LiTS and 3D-IRCADb-01 datasets, demonstrating strong performance in segmenting liver and tumor structures. Results indicate that the proposed method achieves high accuracy and robustness across both datasets, outperforming several existing architectures in capturing complex anatomical boundaries. The use of data augmentation techniques, including rotations and flips, enhances generalization to diverse imaging conditions. Owing to its relatively low parameter count and efficient resource utilization, the model is well-suited for deployment in real-time or resource-constrained clinical environments. These findings highlight the model’s potential as a dependable and computationally efficient tool for automated liver and tumor segmentation in CT imaging, paving the way for broader adoption in clinical workflows and future extension to other anatomical segmentation tasks.
One of the most dangerous life-threatening conditions is brain tumors. It caused in the brain cells and tissues that grow in some abnormal ways. This abnormal grow may be diagnosis as benign (non-cancerous) or malignant (cancerous). The earlier and more accurate diagnosis acts the critical role in proper and ef-fective treatment. There are many ways for brain tumor diagnosis. Magnetic Resonance Imaging (MRI) is widely used for brain tumor detection. As can produce clear, detailed brain images. Recently, deep learning (DL) techniques have shown significant success in aiding radiologists with accurate and rapid tumor clas-sification. To enhance diagnostic accuracy, we proposed a novel hybrid deep learning model in this study. The proposed model combines multiple transfer learning architectures with a Vision Transformer (ViT). Feature extraction is performed using ViT and classification is done by InceptionV3, VGG16, VGG19, ResNet-50, and ResNet-101 on publicly available MRI datasets. Our model can classify categorical-class classification (Glioma, Meningioma, Pituitary Tumor, and No Tumor). Experimental results show that In-ceptionV3 consistently outperforms other models in classification. The proposed model achieves 99.16% accuracy for multi classification and 99.6 for binary classification
Gamification has become a widely adopted strategy in education to enhance student motivation, engagement, and performance. While previous studies confirmed its potential, most relied on short-term measures, self-reported data, or outdated evaluation methods. Other works applied machine learning to predict student outcomes but stopped at classification, without offering interventions to support at-risk learners. As a result, practical assistance for weak or failing students remained limited. This study addresses these gaps through a combined approach that integrates gamification, machine learning, and personalized learning. The research was conducted in two stages. First, datasets from three gamified platforms—Duolingo, Kahoot, and Classcraft—were analyzed using classification algorithms such as J48, Random Forest, KNN, and SVM. The results showed that Kahoot and Classcraft achieved higher accuracy than Duolingo, but none of the platforms provided mechanisms to assist underperforming students. Second, the same machine learning techniques were applied to real student data from the Faculty of Arts. In this phase, a personalized learning model was developed to provide targeted support through adaptive content, motivational feedback, and follow-up sessions. The proposed model achieved 92% accuracy, reduced the failure rate by 22%, and increased student engagement by 15%. By benchmarking against existing platforms and comparing with prior studies, this research demonstrates that combining gamification with personalized, data-driven interventions offers a more effective strategy for supporting struggling learners and improving overall academic outcomes.
Potato, a crucial global crop, contributes significantly to economic growth, job creation, and food security. Rich in carbohydrates, fiber, magnesium, potassium, and vitamin C, it remains vulnerable to devastating blight diseases like early and late blight. Traditional methods for detection prove ineffective. Early and automated detection is paramount to minimize potato yield and quality losses, protecting farmer livelihoods. While Techniques used in DL and ML have been explored for potato blight detection, accuracy and computation time require improvement. This paper proposes a CNN designed to achieve high accuracy using reduced number of parameters and shorter processing time. Utilizing the PlantVillage dataset of 9485 images, to measure the effectiveness of the model, four evaluation metrics were used: precision, recall, F1-score, and accuracy, achieving 99.506%, 99.527%, and 99.515%, 99.65%, respectively. Compared to DL models that were pre-trained and previous work, the presented model outperformed all compared models, attaining an accuracy of 99.65% while boasting only 347,971 trainable parameters and an image processing time of 0.71 seconds.
Blockchain technology is among the technologies that attracts great attention by both academic and business communities. It employs consensus algorithms to address security and decentralization challenges in distributed systems. Among the disadvantages of these algorithms are their high-power consumption, low resistance to Sybil and 51% attacks, and low degree of decentralization. This paper introduces a comparative study of a set of blockchain consensus algorithms previously proposed in literature. The selected algorithms are all based on the Proof-of-Work (PoW) algorithm which forms the core of consensus for many blockchain networks. The algorithms that form the implemented comparative study are PoW, Compute-and-Wait PoW (CW-PoW) and Green-PoW. The three algorithms are compared and evaluated in terms of four different criteria. These criteria include the degree of decentralization among blockchain nodes, block generation time, transaction issue rate, and total computation time. The evaluation results showed that PoW was the best in terms of block generation time; number of transactions issued per second and total computation time. On average, PoW achieved a block generation time of only 0.21 seconds which is much lower than the average block generation times of both CW-PoW and Green-PoW with average block generation times of 143 seconds and 9 seconds, respectively. The average number of transactions issued per second for PoW, CW-PoW and Green-PoW were 14951, 37, and 10921, transactions per second, respectively. The obtained average total computation time for PoW, CW-PoW and Green-PoW were 3, 3109 and 1965 seconds, respectively. Green-PoW showed the best compromise between all evaluated metrics.
Classifying minerals from thin-section microscopic rock images plays a vital role in modern geoscience, supporting tasks ranging from petrographic analysis to mineral exploration and educational applications. While recent advances in machine and deep learning have enabled significant progress in automated image classification, the success of these models depends heavily on the quality and sufficiency of the training datasets. Despite the availability of several geological image datasets, questions remain about whether these datasets are adequate to support robust, generalizable artificial intelligence models in real-world geological contexts. This paper addresses a critical yet underexplored issue in this field, which is the sufficiency of the currently available thin-section image datasets for training reliable classification models. To investigate this issue, a set of widely used machine and deep learning models, namely: KNN, DT, SVM, CNN, and ANN, were applied to selected geological datasets, namely: Igneous and Metamorphic Dataset, Sedimentary Dataset, and Thin Section-1 Dataset, under fair experimental conditions. The results, which ranged from 3.4% to 73.7% accuracy, reveal notable limitations in the generalization performance of these models across different datasets, pointing to issues such as class imbalance, low image quality, and limited class diversity. The findings highlight the urgent need for developing a more comprehensive, balanced, and high-resolution geological dataset that can effectively support automated mineral classification systems. This study contributes to the foundational understanding of dataset sufficiency in geological AI applications and sets the stage for future work in dataset construction and model optimization.
Rapid assessment of disaster-damaged buildings is essential for effective disaster recovery and response. However, the traditional methods of building damage assessment are often labour-intensive and time-consuming. In this study, an ensemble deep learning approach is proposed that combines pre-trained VGG16 and ResNet50 convolutional neural networks (CNNs) for automated classification of building damage. First, each model is fine-tuned individually using an augmented dataset to enhance generalization. The prediction probabilities generated by all base learners are then concatenated to form a new feature set for each image sample. Subsequently, a second-level classifier is constructed using the outputs of these networks, allowing the ensemble method to automatically learn the optimal weighting of each base learner’s contribution. At this stage, Random Forest (RF) classifier and Support Vector Machine (SVM) are used, and their performance is compared with a traditional average ensemble approach which gives equal weights to all basic learners. Proposed models are fine-tuned and tested on the challenging xBD dataset. The individual models achieve accuracies of 88.86% (ResNet50) and 89.29% (VGG16), while the ensemble approach using the SVM classifier achieves the highest accuracy of 90%. These results demonstrate that the ensemble method with RF and SVM classifiers outperforms single CNN models and average ensemble methods, providing a robust and efficient solution for disaster damage assessment.
Deep generative models are being explored as a promising approach for generating distance matrix representations of protein tertiary structures. These models capture the complex spatial arrangement of amino acid residues, but challenges remain in ensuring the generated structures are physically realistic, diverse, and capable of reconstructing missing regions accurately. A key question is whether these models can generate protein structures that align with experimentally determined structures while preserving biological relevance. This review examines recent advancements in generative modeling for protein structures, discussing various methods, their strengths and limitations, and the evaluation metrics used. Recent studies have demonstrated significant improvements in model performance by integrating biophysical constraints, multi-scale representations, and hybrid learning strategies, offering promising directions for enhancing structure prediction accuracy, native-like structural integrity, and biological relevance. It also highlights the need for optimizing model architectures and improving validation strategies to enhance protein structure prediction and expand their role in structural biology.
Hyperspectral imaging provides fine-grained spectral information for land cover classification. Still, it faces critical challenges, including high dimensionality, spectral redundancy, and limited labeled samples, which can lead to overfitting in deep learning models. This paper introduces a unified framework that synergistically integrates Linear Discriminant Analysis with Swin Transformers, engineered explicitly for robust hyperspectral classification under data-scarce conditions. The framework employs systematic Bayesian optimization to automatically determine optimal parameters, while LDA serves as a class-discriminative feature compressor that maximizes inter-class separability before spatial modeling. The Swin Transformer then captures hierarchical spatial-spectral dependencies via shifted window attention. Extensive experiments on Indian Pines, Pavia University, and Salinas datasets demonstrate state-of-the-art performance, with statistically validated accuracies of 97.82%, 99.72%, and 99.25% respectively. The framework maintains >92% accuracy with only 10% labeled data and achieves a 38% reduction in computational requirements, highlighting its practical efficiency. Comprehensive ablation studies and statistical significance testing confirm the individual and synergistic contributions of both components. These results establish a new paradigm for combining classical dimensionality reduction with modern attention mechanisms, offering significant potential for real-world remote sensing applications in precision agriculture and environmental monitoring.
Electronic Health Records (EHR) have become a cornerstone of modern healthcare, enabling efficient storage, retrieval, and management of patient data. With the rise of IoT devices such as smartwatches, fitness trackers, and other wearable technology, accessing EHR in real-time has become increasingly common. These devices often operate in constrained environments with limited computational resources, necessitating cryptographic algorithms that are both lightweight and secure. This research evaluates some lightweight cryptographic algorithms—Speck, AES, Simon, LEA, DES, and RSA—across four operating systems with different resources. The study focuses on key performance metrics, including execution time, energy consumption, memory usage, and CPU usage, to determine the best algorithm for securing EHR on resource-constrained IoT devices. The findings highlight Speck and AES as the most efficient algorithms, providing quick, secure, and resource-friendly operations. In contrast, DES and RSA exhibited high resource demands, making them unsuitable for such applications. This study underscores the need for carefully selecting cryptographic algorithms to enhance security and efficiency in IoT-based healthcare systems.
