Distributed Denial-of-Service (DDoS) attacks remain among the most disruptive network threats, and detectors that generalize across attack families with low false-alarm rates are still an open problem. Propose an adaptive hybrid ensemble that unifies two gradient-boosting learners (Random Forest and Gradient Boosting) with three deep neural base learners (DNN, CNN-1D, and LSTM) under a weighted soft-voting rule whose weights are produced by a Reinforcement Learning (RL) policy. The RL agent treats the ensemble-weight simplex as its action space, observes a state vector built from validation-set diagnostic statistics, and is trained by REINFORCE-with-baseline to maximize a reward equal to validation F1 minus a small calibration penalty. The framework is formalized as a Markov decision process with one stochastic step per training episode, which decouples ensemble-weight learning from the (non-differentiable) outer F1 objective. On a 10,000-sample, 25-feature, five-class benchmark with 7% label noise, the proposed system reaches weighted F1 = 0.846, accuracy = 84.7%, MCC = 0.781, AUC = 0.952, and ECE = 0.039. Friedman and Nemenyi post-hoc tests over 50 CV folds confirm the RL-guided ensemble is significantly better than every individual base learner and uniform voting at α = 0.05 (Cohen's d = 0.96). An ablation isolates the RL policy and gradient boosting as the main drivers; a label-noise robustness study shows graceful degradation up to 20%; a head-to-head comparison against the Bonobo Optimizer (BO), GA, PSO, GWO, and WOA shows the best F1/wallclock trade-off.
Pneumonia detection through medical imaging, especially using CT scans or X-rays, presents notable challenges due to the subtle and often unclear signs of the disease. This paper introduces a novel neural network model, the Compact Convolutional Transformer (CCT), designed to address these challenges by optimizing detection accuracy. The CCT model incorporates configuration dropout in its convolutional layers to enhance both robustness and precision.Experiments conducted on a dataset of 5,856 chest X-ray images from pediatric patients aged one to five years demonstrated the model's effectiveness, achieving a remarkable 97% accuracy, 97% recall, 98% precision, and an F1-score of 98%. When compared to state-of-the-art models like DarkNet-53 and VGG-19 + GradCAM, which achieved F1-scores of 97.3% and 95.61% respectively, the CCT model consistently matched or outperformed them, particularly when dealing with smaller and more complex datasets. Even models such as CNN + Bayesian Network, which used larger datasets, only reached an F1-score of 96.3%.These results underscore the superior efficiency and accuracy of the CCT model, highlighting its potential for broader applications in medical diagnostics and image analysis, especially in pneumonia detection.
The human brain plays a vital role in regulating bodily functions, and abnormal cell growth may lead to life-threatening brain tumors. Automated computer-aided diagnosis systems are therefore essential to support early detection from MRI images. This study investigates brain tumor classification using Gray Level Co-occurrence Matrix (GLCM) feature extraction combined with Support Vector Machine (SVM) classification. Unlike prior works that typically employ a single kernel configuration, this study conducts a systematic comparison of four SVM kernels: linear, polynomial, radial basis function (RBF), and sigmoid under a consistent preprocessing pipeline and structured hyperparameter tuning framework. GLCM features, including energy, contrast, correlation, and homogeneity, were extracted at multiple distances and angles. Kernel performance was evaluated using controlled hyperparameter search procedures to ensure fair comparison across models. Experimental results on a binary MRI dataset consisting of 2,800 images demonstrate that the RBF kernel achieved the highest accuracy of 96% with C = 100 and gamma = 10, outperforming polynomial (74%), linear (72%), and sigmoid (71%) kernels. The findings highlight the importance of systematic kernel evaluation and parameter sensitivity analysis in texture-based medical image classification. The proposed GLCM–SVM framework provides a computationally efficient and interpretable approach that may support preliminary decision-aid systems for brain tumor screening.
Mental health severity assessment is often hindered by limited access to professional services and the time required for clinical evaluation. This study proposes an interpretable soft voting method to classify the severity levels of depression, anxiety, and stress using DASS-42 questionnaire data. The proposed framework integrates Logistic Regression, Random Forest, Support Vector Machine, and Extreme Gradient Boosting, and is evaluated on 35,445 anonymized responses from a public psychometric dataset. Model performance was assessed using accuracy, precision, recall, F1-score, macro-averaged and weighted F1-score and the precision-recall curve under stratified cross-validation for class imbalance between the normal and mild classes. Explainable Artificial Intelligence using SHAP was employed to interpret model decisions. The soft voting achieved strong predictive performance, with accuracy values of 0.98 for depression, 0.99 for anxiety, and 0.97 for stress, outperforming or matching individual base models. SHAP analysis identified clinically consistent features contributing to model predictions; due to computational constraints, SHAP analysis was not applied to the SVM model. Despite strong performance, the use of secondary self-reported data and class imbalance, particularly the underrepresentation of normal and mild cases, were limitations. The proposed model demonstrates the potential of interpretable soft voting as a decision-support tool for mental health severity stratification in resource-constrained settings.
Ulos is a traditional Batak textile with high cultural significance, yet artisans lack tools to modernize designs while preserving heritage. Previous research primarily uses single-objective methods that fail to accommodate user preferences. This study introduces the first multi-objective framework for traditional textile coloring, differentiating itself by using a large language model (LLM) as a natural language interface for non-technical weavers. The LLM translates user design preferences into dynamic objective functions to guide a Non-dominated Sorting Differential Evolution (NSDE) algorithm. Performance is assessed using the epsilon indicator, aesthetic metrics (contrast and colorfulness), and user preference scores. The system achieved high convergence with an epsilon-indicator value and user preference scores reaching 0.66. Additionally, a paired t-test (p=0.07) demonstrated the algorithm's robustness across parameter variations. This approach enables artisans to create culturally authentic, aesthetically optimized designs through intuitive interaction.
Indonesia has a high diversity of medicinal plants that are widely used in traditional healthcare practices. Identification of medicinal plants is commonly based on leaf morphology; however, similarities in leaf shape, texture, and color often cause misidentification, particularly among non-experts. This limitation highlights the need for an automated and reliable identification approach. The primary objective of this study is to develop and evaluate a deep learning–based system for the automatic identification of medicinal plants using leaf images, with a specific focus on comparing the performance and efficiency of MobileNetV2 and ResNet50V2 architectures. The research design adopts an experimental approach using an internally collected dataset of medicinal plant leaf images representing multiple plant classes. The dataset is divided into training and testing sets to evaluate model generalization. The methodology involves image preprocessing steps, including resizing, normalization, and data augmentation, followed by the application of transfer learning using MobileNetV2 and ResNet50V2 as feature extractors. Both models are trained under the same experimental settings and evaluated using standard classification metrics, including accuracy, precision, recall, F1-score, and confusion matrix analysis. The main outcomes and results indicate that both deep learning models achieve high classification performance. MobileNetV2 achieves an accuracy of 98.77%, precision of 98.84%, recall of 98.77%, and F1-score of 98.77%, while ResNet50V2 achieves an accuracy of 97.53%, precision of 97.87%, recall of 97.53%, and F1-score of 97.58%. The results demonstrate that MobileNetV2 provides slightly superior performance with lower computational complexity. In conclusion, lightweight deep learning architectures such as MobileNetV2 are effective and efficient for medicinal plant leaf identification and are suitable for implementation in mobile or resource-constrained environments.
Early detection of plant leaf diseases is critical for minimizing crop losses and supporting precision agriculture. While Convolutional Neural Networks (CNNs) have demonstrated high accuracy in image-based diagnosis, conventional architectures may not optimally balance spatial localization and channel-wise feature refinement, particularly in multi-crop classification settings. This study proposes a redundancy-aware dual-attention architecture, termed ATSA-DenseNet, which integrates the spatial branch of the Convolutional Block Attention Module (CBAM-Spatial) with Efficient Channel Attention (ECA) within a DenseNet121 backbone. Unlike prior dual-attention frameworks that retain full CBAM and introduce channel-level redundancy, the proposed design isolates complementary spatial and channel mechanisms to improve representational efficiency without increasing computational complexity. The framework is evaluated on controlled multi-crop PlantVillage-derived datasets comprising tomato, potato, pepper, and maize. Across both 3-crop and 4-crop configuration, ATSA-DenseNet consistently outperforms baseline DenseNet121 and single-attention variants, achieving 99.94% accuracy and 0.9994 macro-F1 on the 4-crop setting while maintaining a lightweight footprint (6.96M parameters, 2.87G FLOPs). Grad-CAM visualizations indicate improved localization of disease-relevant regions compared to the baseline. While results are obtained under controlled imaging conditions, the findings demonstrate that redundancy-aware dual-attention enhances feature discrimination efficiency in multi-class agricultural classification tasks. Future work will extend validation to real-field datasets with natural variability.
Translating natural language questions into MongoDB queries is critical for flexible data access in current NoSQL systems. However, semantic ambiguity in user questions and MongoDB's dynamic schema make this work difficult. This study presents QMQL (Question to Mongo Query Language), a hybrid approach meant to address these challenges. QMQL combines a Graph Attention Network (GAT) for refining schema elements with a Retrieval-Augmented Generation (RAG) mechanism that employs BERT embeddings to retrieve relevant schema and resolve semantic ambiguity. A T5-base model is used to generate a MongoDB query corresponding to the user’s question. An experimental evaluation on an extended dataset encompassing various real-world domains demonstrates the effectiveness of the proposed approach. QMQL achieves excellent performance with an EMA of 0.89, an EM of 0.91, and a BLEU score of 0.95, exceeding previous approaches, particularly for semantically ambiguous questions and sophisticated queries across flexible MongoDB schemas,
The 5G mobile networks offer extra benefits in terms of high data rates, lower latency, and more coverage in comparison to 4 G networks. However, the 5G network offers new levels of data transfer and processing speeds, ensuring users do not disconnect as they move from one cell to another. By considering these issues, this paper proposes a new blockchain-based, scalable, and reliable 5G handoff authentication system. The proposed approach is intended to ensure authentication using Exponential Elliptic Curve-Assisted Encryption (EEE) between the user and the base station. After the successful authentication, the user stores the details in the source base station. In a blockchain-based 5G handoff authentication system, when a user device enters the range of a new base station, it sends a handoff request using a pseudorandom frequency-hopping sequence. The serving base station sends a handover command to the mobile device, containing details about the target base station and the next frequency in the hopping sequence. This request is disseminated through an improved gossip algorithm that minimizes communication overhead and accelerates node authentication in the blockchain for consensus and validation. Similarly, in blockchain networks, the gossip protocol ensures that every node in the network receives information about the request for handoff while ensuring that the messages sent are not redundant. Therefore, the joint graph-based Delegated Practical Byzantine Fault Tolerance (JG-DPBFT) consensus process is utilized to verify the handoff.
Fine-grained surgical action recognition in laparoscopic videos remains a challenge, even with recent advances in deep learning. While current VideoMAE approaches reach 89.11% accuracy on cholecystectomy tasks, they face specific limitations. Random masking strategies often miss surgical instruments that occupy only 10% to 15% of frames. Furthermore, context-independent models struggle with visually similar actions across different phases, and symmetric two-stream architectures tend to waste computational resources. To solve this, we developed SA-VideoMAE, a surgical-aware video masked autoencoder specifically designed for laparoscopic action recognition. Our method utilizes surgical-aware adaptive masking that integrates YOLOv7x object detection to prioritize instrument patches. This increased instrument visibility from 10% to 60% during training, ensuring the model focuses on action-relevant regions rather than static backgrounds. We also utilized phase-conditioned hierarchical attention to inject learnable phase embeddings into the attention mechanisms, enabling the model to disambiguate visually similar actions based on surgical context. For efficiency, our asymmetric dual-stream architecture processes RGB using ViT-Base (86M parameters) and optical flow using ViT-Tiny (5.7M parameters), achieving a 47% parameter reduction compared to symmetric designs. Our training process then balanced reconstruction, classification, temporal consistency, and phase prediction through a novel multi-objective optimization strategy. Results from Cholec80's Calot's Triangle Dissection phase show 93.5% accuracy, representing a 4.4 percentage-point improvement over the verified baseline. Notably, challenging action recall improved from 51% to 74% while maintaining real-time inference at 62ms per clip. These findings demonstrate that encoding surgical domain knowledge into video architectures significantly enhances action recognition performance.
Deep learning (DL) has a significant impact on X-ray images for diagnosing and categorizing a range of lung disorders. The proliferation of extensive annotated image datasets has resulted in the emergence of convolutional neural networks (CNNs) as a useful instrument for the tasks of image recognition and categorization. Despite this abundance, the primary challenge in medical diagnosis continues to be the classification of these images. This research aims to enhance image classifiers by using the CNN model; both training and testing datasets underwent analysis using the suggested CNN system. An analysis and comparison are conducted on the impact of feature extraction using the Principal Components Analysis (PCA) technique. The study attains maximal classification efficiency by preparing images through dimensionality reduction before classification, concurrently enhancing the efficacy of CNNs in feature extraction. The optimal tuning strategies for enhancing the performance of the proposed CNN were found to include boosting the number of epochs, changing the optimizer, and decreasing the learning rate, as well as improving algorithm gains by using pre-trained weights. The suggested system outperforms previously utilized approaches like VGG or DenseNet, with more than 99.80 percent accuracy, precision, recall, and F1-score values. The suggested methodology demonstrates considerable promise for enhancing the efficacy and precision of lung disease diagnoses derived from chest X-ray images, thereby offering clinicians beneficial decision support and accelerating the implementation of treatment strategies. Furthermore, the developed model facilitates the identification of pulmonary ailments, encompassing critical conditions like COVID-19, thereby facilitating timely and efficacious patient care.
In conventional tuberculosis diagnosis, 100-300 fields of view (FOVs) must be observed, which can lead to observer fatigue. For both tasks, an automatic stitching framework was developed that extends conventional feature-based transformations by incorporating affine-geometry-based feature matching and RANSAC-based homography refinement, thereby accounting for the unique low-texture morphology and irregular patterns of Mycobacterium tuberculosis in ZN-stained sputum smears. The system was tested on a set of 10 overlapping image pairs with a fixed overlap of 30%. Among the evaluated image pairs, the proposed optimized method achieved a 100% success rate. Objective zero-pixel metric-based quantitative analysis also validated higher transparency quality compared to other methods. The proposed SURF implementation reached a minimum number of 345.263 zero-pixels, outperforming standard SURF (964.247) and SIFT (1.069.687). This improved robustness to rotation and illumination variations made the optimized SURF-affine framework a preferred choice for automatic TB diagnosis systems.
The Job Shop Scheduling Problem (JSSP) is a combinatorial optimization problem that is NP-hard and highly complex, particularly in modern manufacturing environments associated with industry. Conventional metaheuristic methods such as Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) are capable of generating solutions in a relatively short time; however, they often face limitations in solution quality due to premature convergence and limited adaptability to dynamic problem conditions. In contrast, deep learning approaches such as Sequence-to-Sequence (Seq2Seq) offer strong representational capabilities for modeling operation sequences, although they still encounter challenges related to training stability and generalization. This study proposes a hybrid approach that integrates a Seq2Seq encoder–decoder architecture with an adaptive metaheuristic mechanism to enhance scheduling optimization performance. The Seq2Seq model is utilized to learn underlying patterns in operation sequences, while the adaptive mechanism dynamically adjusts search parameters based on makespan evaluation. The experiments are conducted using datasets from the OR-Library, specifically the 10×10 and 15×15 scenarios, to evaluate the performance and scalability of the proposed method. The experimental results demonstrate that the Seq2Seq + adaptive metaheuristic approach consistently produces lower makespan values compared to GA and PSO. For the 10×10 dataset, the proposed method achieves a makespan of 932, outperforming GA (1095) and PSO (1047). Similarly, for the 15×15 dataset, it attains a makespan of 1050, which is better than GA (1250) and PSO (1200). Although the proposed approach requires slightly longer computational time, the improvement in solution quality indicates that it effectively balances exploration and exploitation.
Accurate breast ultrasound (BUS) lesion segmentation is critical for early diagnosis but is challenged by image artifacts and the reliance of foundation models on manual prompting. Existing automated frameworks often lack robust fail-safe mechanisms, leading to missed diagnoses. To address this reliability gap, this study proposes a novel, fully automated hybrid segmentation framework that synergistically integrates three key components: (1) a recall-optimized YOLOv9 detector tailored to minimize clinical false negatives; (2) a MedSAM2 foundation model efficiently fine-tuned via Low-Rank Adaptation (LoRA) for ultrasound specifics; and (3) a statistical fallback mechanism that acts as a crucial safety net to recover spatial prompts during detection failures. Evaluated on the public BUSI dataset, the recall-dominant detection module achieved a Recall of 0.8238. Supported by this robust prompting and fallback strategy, the segmentation module achieved a Dice coefficient of 0.8818 and an IoU of 0.8113. By effectively integrating specialized detection with adaptive segmentation and a statistical fail-safe, the proposed pipeline offers a highly reliable automated approach for computer-aided screening systems.
The growing volume of digital text content requires automated summarization systems as fundamental tools that enable users to access information at high speed. Automatic Multi-Document Summarization (MDS) requires systems to produce a summary that combines essential information from multiple documents. The extractive methods, which rely on lexical signals and sentence-based rules, yield only repetitive results because they cannot capture complex thematic relationships. This study developed an improved extractive MDS model that combines topic modeling with pattern-based semantic indicators and a method to choose diverse sentences. The model employs LDA to identify concealed thematic structures, retrieves typical word patterns to improve topic models, and selects topics via a greedy algorithm that reduces redundancy to achieve appropriate salience and coverage. The proposed system achieves better results than classical baselines on the DUC 2006 and DUC 2007 datasets, outperforming Lead, CLASSY04, KL-SUM, LexRank, TextRank, and PETMSUM. The system demonstrates superior performance over all baseline methods, achieving better results on the ROUGE-1, ROUGE-2, and ROUGE-SU4 evaluation metrics. The results show that extractive summarization tasks achieve their best performance when topic–pattern representations are combined with diversity-aware scoring methods.
Cyberattacks are becoming increasingly sophisticated, necessitating defense mechanisms that go beyond simple detection to include severity assessment for prioritizing mitigation. This study proposes a comprehensive machine learning framework to classify cyberattack severity levels (Low, Medium, High) using a modern, high-dimensional dataset. Addressing the critical challenge of class imbalance, the research integrates the Synthetic Minority Oversampling Technique (SMOTE) with a rigorous feature selection process involving SelectKBest. Four algorithms Naive Bayes, K-Nearest Neighbor (KNN), Random Forest (RF), and Support Vector Machine (SVM) were evaluated using 10-fold cross-validation. The results demonstrate that the SVM model with an RBF kernel achieves superior performance with an accuracy of 97.30% and a False Negative Rate (FNR) of only 3.1% for high-severity threats. This research contributes a robust, data-driven approach to severity classification that effectively handles feature non-linearity and class imbalance, offering actionable insights for real-time security operations.
Streaming anomaly detection is a difficult task because the data distribution is changing, and concept drift degrades the performance of traditional detection methods. To solve the issue, this article suggests the HAD-CDA (Hybrid Anomaly Detection with Concept Drift Adaptation), which is a combined system that can detect local anomalies in the time series with LSTM Autoencoders (LSTM-AE) and global anomalies in the distribution with Quant Tree-EWMA (QT-EWMA). The proposed framework accomplishes three objectives: (i) the use of the dynamic weighting mechanism, which automatically changes the contribution of each component (lambda between 0.2 and 0.8) according to their effectiveness, (ii) two concept drift detectors are proposed, i.e., the Kolmogorov-Smirnov and Page-Hinkley tests, to allow detecting concept drift; and (iii) the use of Elastic Weight Consolidation (EWC) to reduce catastrophic forgetting during update of the model. Experiments on four real-world streaming datasets, HTTP, SMTP, ForestCover, and Shuttle, indicate that HAD-CDA has AUCs of 0.95-0.97, an 8-9% improvement over state-of-the-art methods, and F1-scores of 0.81-0.95. The recall measure obtained by the LSTM-AE element is 0.8094, compared to 0.9399, and the specificity of QT-EWMA is very high at 0.9399. The framework is highly adaptable to different types of drifts, regaining around 9092 performance levels before drift, and 1525 windows would be needed by a baseline method, with a low processing latency of 12.430.9 ms per window. Having a per-sequence complexity of O(1), memory cost of O(n), low DIS (approximately 0.08), and stability indices of 0.02-0.04, the suggested HAD-CDA framework is an accurate, efficient, and robust solution to real-world streaming anomaly detection in changing data
Accurate weather prediction is essential for supporting various human activities and mitigating the impacts of changing atmospheric conditions. Recent advances in artificial intelligence have enabled the development of data-driven forecasting models capable of capturing complex relationships among meteorological variables. This study proposes an Artificial Neural Network (ANN)-based weather prediction model using multi-sensor weather data, including temperature, humidity, precipitation, solar irradiance, and wind velocity. The proposed ANN architecture consists of an input layer, two hidden layers, and an output layer. Model performance was evaluated using three training–testing data splits (90/10, 80/20, and 70/30) with 100 and 150 training epochs. Prediction performance was assessed using accuracy and Root Mean Squared Error (RMSE). Experimental results demonstrate that the proposed model achieves the best performance with a 70/30 training–testing split and 150 training epochs, providing the highest prediction accuracy and the lowest RMSE among the evaluated configurations. These findings indicate that a relatively simple ANN architecture can effectively model multi-sensor weather data and provide reliable weather predictions.
In this paper, we developed an effective wrapper-based model to optimize the recognition of physical human daily life activities recorded by smartphone built-in sensors. The proposed model employs Sequential Forward Selection (SFS) method in combination with K-Nearest Neighbor (KNN) classifier based on a group of five commonly used distance measures: (Manhattan, Euclidean, Chebyshev, Canberra, and Correlation) each having a specific geometric neighborhood interpretation. The proposed SFS-KNN multi-distance approach enables each distance metric to guide the feature selection process. It examines how the resulting feature subsets, determined by each distance, impact the overall recognition performance. The goal is to identify the best feature subset that achieves the highest accuracy with the lowest dimensionality. . Our proposed distance-based wrapper model was validated on two publicly available WISDM and UCI-HAR datasets under 10-fold cross-validation. The experimental results obtained on both datasets showed that the performance of the proposed model is significantly affected by the distance measure used due to generating different feature subsets for each distance. For the WISDM dataset, our model achieves an overall accuracy of 93.39% based on Euclidean distance, with a reduction ratio of 71.42%. It also offers a substantial reduction of 95.9% in the feature dimensions of the UCI-HAR dataset using Correlation distance, with a recognition rate of 99.33%. These outcomes confirm the superiority of our wrapper model over other feature selection-based approaches proposed for the same datasets.
The prevalence of brain tumors has been increasing annually, and headaches, a common initial symptom, represent the most common manifestation. However, there is a paucity of research on effective methods of assessing brain tumors. This study proposes a novel approach by introducing various modality fusion techniques based on their fusion levels, which are then categorized into four groups: single-modal, data-level fusion, feature-level fusion, and multilevel fusion. A total of 51 combinations are designed to evaluate the efficacy of these fusion techniques and modality configurations. The experiments used a BraTS2021, which comprises four magnetic resonance imaging (MRI) sequences (flair, t1, t1ce and t2). Initially, the image was pre-processed, encompassing data selection, conversion, and normalization. Subsequently, it was input into a 13-layer CNN architecture for feature extraction. Classification was facilitated by a soft voting method in ensemble learning, incorporating support vector machine (SVM), k-nearest neighbor (KNN), logistic regression, random forest, and decision tree algorithms. The predictive efficacy of the model was rigorously assessed through a comprehensive suite of metrics, prominently featuring accuracy, AUCROC, AUCPR, Cohen's Kappa, and MCC. The results indicate that multilevel fusion exhibits optimal performance, with an average accuracy of 95.84%, followed by feature-level fusion and data-level fusion, at 95.12% and 94.77%, respectively. The optimal fusion technique was identified as the combination with the FF configuration (1,2),3,4), producing an accuracy of 96.62%. The best-model combination proposed exhibited an accuracy difference of nearly 6% from the baseline model, underscoring the efficacy of the proposed approach. These empirical results establish a robust baseline for future investigations into sophisticated fusion architectures across hierarchical integration levels.
