Kinetik: Game Technology, Information System, Computer Network, Computing, Electronics, and Control

Hybrid Deep Embedding and LightGBM Meta-Classification for Multi-Class Intrusion Detection in Wireless Sensor Networks

Ridha Muldina — 2026-10-25

Internet of Things (IoT) applications require high service availability, and Wireless Sensor Networks (WSNs) are increasingly used in such applications, but WSNs are subject to Denial of Service (DoS) attacks that drain resources and interrupt communication. However, in such scenarios, the effectiveness of machine-learning-based Intrusion Detection Systems (IDS) was typically compromised by excessive class imbalance and inter-class overlap, resulting in high accuracy but low sensitivity to minority attack classes. In this research, we present a Hybrid Deep Embedding and LightGBM Meta-Classification system to improve multi-class IDS performance on the WSN-DS dataset. The contribution of each part was quantitatively evaluated using Stratified 5-Fold Cross-Validation in a five-stage experimental design: models, data balancing, loss/weighting strategies, optimal combination, and the suggested system. The complementary metric was reported as accuracy, and minority-aware metrics were applied, notably Macro-F1, balanced accuracy, and Matthews Correlation Coefficient (MCC). The results show that the independent use of data balancing and loss adjustment does not consistently improve performance in inter-class overlap scenarios. Conversely, the suggested system that learned deep embedding representations and modified the decision boundaries using a LightGBM meta-classifier demonstrated more consistent performance, with the best configuration achieving Macro-F1 of 0.977, balanced accuracy of 0.975 and MCC of 0.981. These results revealed that the improvement of representation separability and decision boundary flexibility was more consistent than the separate treatment of data imbalance in improving the performance of multi-class IDS.

Robust Video Encryption using Multi-Chaotic Map Cascade to Mitigate Statistical and Differential Attacks

Heru Lestiawan — 2026-10-05

The rapid proliferation of digital video transmission has necessitated robust security mechanisms to safeguard sensitive visual content against unauthorized access and interception. Conventional cryptographic algorithms, however, often struggle with the efficiency demands of high-volume video data. This study established and validated a lightweight yet highly secure video encryption scheme utilizing a Multi-Chaotic Map Cascade architecture. The proposed approach integrated the SHA-256 hash function for dynamic key generation with a hybrid hyperchaotic system comprising Henon, Lorenz, and Tent maps. This configuration facilitated a rigorous encryption process involving pixel coordinate permutation and bidirectional value diffusion. Comprehensive experimental evaluations were conducted on standard video sequences ranging from QCIF to Full HD resolutions to assess consistency and scalability. The results demonstrated that the algorithm achieved near-perfect randomness, with an average information entropy of 7.9988 bits and a Number of Pixels Change Rate of 100 percent, effectively satisfying the strict avalanche criterion. Furthermore, histogram analysis confirmed a statistically uniform distribution validated by Chi-Square tests. The scheme also exhibited strong resistance against visual reconstruction, maintaining Peak Signal-to-Noise Ratio values below 10 dB for encrypted content. Additionally, Structural Similarity Index (SSIM) analysis confirmed near-zero structural correlation (averaging 0.018) in encrypted frames while achieving perfect reconstruction (SSIM = 1.0) during decryption. Adjacent pixel correlation analysis demonstrated exceptional randomness with correlation coefficients approaching zero (|r| < 0.006) across horizontal, vertical, and diagonal directions, effectively eliminating spatial dependencies.

Evaluating CLAHE and Temporal Smoothing Impact on Deep Learning-Based Young Crescent Moon Video Detection

Bayu Krisna Murti — 2026-10-05

Detecting the young crescent moon from video data poses significant challenges due to low contrast against the twilight sky, sensor noise, and atmospheric interference. Frame-wise contrast enhancement, such as Contrast Limited Adaptive Histogram Equalization (CLAHE), can introduce inter-frame intensity fluctuations that degrade both visual quality and the consistency of automated detection. This study evaluates the impact of CLAHE and Temporally Aware CLAHE (TA-CLAHE) preprocessing on YOLOv8n-based young crescent moon detection across three video sequences representing varying observational conditions. Temporal stability was assessed using Flicker Index (FI), Percent Flicker (PF), Temporal Standard Deviation (T-STD), and Frame Difference Mean (FDM). At the same time, detection performance was measured using Detection Consistency Rate (DCR) and centroid stability relative to a fixed ground truth. CLAHE substantially improves detection consistency, raising DCR by up to 16 percentage points over the Raw baseline (Video 1: 60.8% to 76.8%). TA-CLAHE further elevates detection to near-complete coverage — 99.9% on Video 1 and 99.2% on Video 3 — surpassing standard CLAHE by 23.1 and 9.5 percentage points, respectively. Temporal stability also improves: TA-CLAHE reduces FI by 19–32% and T-STD by 20–26%; the reduction in PF is modest (at most 27%) and is retained only as a secondary indicator given its known measurement limitations. These gains involve trade-offs: enhancement slightly loosens per-box localization, and under cloud occlusion, TA-CLAHE raises the invalid detection rate to 17%. Based on these findings, CLAHE is recommended as a sound, low-cost default preprocessing, while TA-CLAHE is preferred for applications requiring an uninterrupted, low-jitter detection track, provided it is paired with an explicit occlusion gate.

An Optimized Hybrid Kernel Support Vector Machine Framework for Multi-Class Brain Tumor MRI Classification

Petrus Sokibi — 2026-10-06

Brain tumor remains one of the most life-threatening diseases worldwide, with early detection and accurate classification still posing major clinical challenges due to complex tissue structures and overlapping visual characteristics in MRI scans. To address this issue, this study proposes a Hybrid Kernel Support Vector Machine (HK-SVM) framework that integrates multiple kernel functions to improve classification performance and tumor localization. The hybridization of nonlinear and linear kernels enhances the model’s ability to capture diverse spatial and textural features, enabling more accurate separation of glioma, meningioma, and pituitary tumor classes. Experimental results show that the Sigmoid and RBF hybrid kernel achieved the best performance, reaching 99.3% accuracy along with precision, recall, and F1 score values of 99%, outperforming other kernel combinations. In addition, a region-based segmentation process was applied to visualize the detected tumor area, demonstrating close alignment with ground truth masks. The proposed method contributes to developing a more interpretable and computationally efficient MRI-based diagnostic framework, offering a reliable alternative to deep learning approaches for medical image classification and paving the way for future integration with deep models such as U-Net or transfer learning CNNs for enhanced segmentation and feature extraction.

Optimizing Non-Playable Characters Navigation through the A-Star Algorithm in Complex 3D Game Environments

Muttabik Fathul Lathief — 2026-10-18

This study presents the implementation and evaluation of the A-Star algorithm for adaptive navigation of Non-Playable Characters (NPCs) in three-dimensional game environments. The research addresses the need for intelligent pathfinding systems that enhance player engagement through realistic NPC behavior and dynamic environmental adaptation. A comprehensive experimental framework compared A-Star against Dijkstra's algorithm and Unity NavMesh across four maze environments with varying complexity and terrain penalty systems. The implementation employs a grid-based environment model with 8-connected neighborhood topology, integrating terrain penalty systems for intelligent route selection. Performance evaluation utilized six metrics: execution time, path quality, waypoint count, total path length, character traversal time, and navigation success rate. Testing involved 10 iterations per algorithm per environment, totaling 120 experimental trials. Results demonstrate A-Star's superior computational efficiency, achieving the fastest average execution time of 0.01272 seconds compared to Dijkstra (0.01775 seconds) and Unity NavMesh (0.02055 seconds), representing 28.3% and 38.1% performance improvements respectively. A-Star exhibited unique intelligent adaptation, selecting alternative routes to avoid high-penalty terrain in 30% of test cases, while other algorithms showed no adaptive behavior. All algorithms achieved 100% pathfinding success rates. Performance consistency analysis revealed trade-offs between speed and stability. A-Star demonstrated higher variability (CV = 33.45%) reflecting adaptive computational behavior, while Unity NavMesh exhibited greater consistency (CV = 23.5%). The findings validate A-Star as an optimal solution for modern games requiring adaptive NPC behavior, contributing to intelligent navigation systems that enhance gameplay dynamics.

An Improved ResNet-50 Based on Squeeze and Excitation for Lung Tuberculosis Classification

Lakui Johary — 2026-10-25

Tuberculosis remains a major global health problem due to its high mortality rate and the difficulty of early diagnosis from medical imaging. Computed tomography (CT-scan) provides detailed visualization of lung structures; however, distinguishing between normal and tuberculosis patterns is challenging due to subtle and complex features. In this study, an improved deep learning model based on ResNet-50 integrated with a Squeeze-and-Excitation (SE) mechanism was proposed to enhance classification performance. A private CT-scan dataset collected from Klinik Utama Centrum, Semarang, was utilized, consisting of imbalanced classes that were addressed through data augmentation applied specifically to the tuberculosis class. The images were resized to 224 × 224 pixels and divided into training and testing sets with an 80:20 ratio. The model was trained using transfer learning with the Adam optimizer and evaluated using confusion matrix-based metrics. The experimental results showed that the proposed model achieved an accuracy of 94.55% and a macro F1-score of 0.9545, outperforming the standard ResNet-50 model. The confusion matrix analysis indicated a significant reduction in misclassification, particularly in tuberculosis detection. Furthermore, single-prediction testing on newly collected 2026 data demonstrated perfect classification results without any misclassification, which were clinically validated by a CT-scan specialist. These findings confirm that the integration of the SE mechanism improves feature representation and enhances model robustness, making it a reliable approach for automated tuberculosis classification from CT-scan images.

A Digital Twin–Based Traffic Control System for Heterogeneous AMRs Using Type-2 Fuzzy Logic

Muhammad Fauzan Hernadi — 2026-10-18

The simultaneous operation of multi-brand Autonomous Mobile Robots (AMRs) introduced interoperability challenges that could lead to path conflicts and collisions between AMRs operating within the same industrial environment. This study proposed a Type-2 Fuzzy logic-based control system integrated with a Digital Twin framework to coordinate the movement of two heterogeneous AMRs, the Hikrobot and the AutoXing units, and to enhance operational safety during simultaneous operation. Unlike prior single-brand or simulation-only approaches, the proposed framework combined real-time Digital Twin monitoring with Type-2 Fuzzy Logic to manage uncertainty in inter-AMR distance and speed data, enabling stop-and-go decisions for the Hikrobot AMR while validating system behavior in a virtual environment prior to physical deployment. Experimental results showed that the baseline method produced 2 inter-AMR collisions, whereas both the Type-2 and Type-1 Fuzzy control systems eliminated all collisions and deadlock conflicts across the tested scenarios. However, this improvement introduced a trade-off in travel time: the baseline method averaged 57 seconds, the Type-1 Fuzzy system averaged 96 seconds, and the Type-2 Fuzzy system averaged 79 seconds, approximately 18% faster than Type-1 despite the added computational complexity of type-reduction. The Type-2 system's response latency was also 0.2 seconds higher than that of Type-1, a difference that remained within a safe operational margin. Overall, the Type-2 Fuzzy system demonstrated superior control decision stability under dynamic conditions with only a marginal latency cost, indicating that the proposed Digital Twin-based Type-2 Fuzzy framework offered a more balanced trade-off between safety and responsiveness than the Type-1 approach.

An Effective IoT Attack Detection System With A Symmetrical Uncertainty Factor and Random Forest

Kurniabudi — 2026-10-13

The rapid expansion of the Internet of Things (IoT) ecosystem has exposed critical security flaws, primarily because the high-dimensional, heavily imbalanced characteristics of IoT network traffic severely compromise existing intrusion detection systems (IDSs). A major limitation in current research is the decoupling of feature selection and classification, which leads to feature redundancy, high computational demands, and less-than-optimal detection rates. To overcome this limitation, we introduce an enhanced IoT attack detection architecture that seamlessly couples the Symmetrical Uncertainty (SU) filter for feature selection with the Random Forest (RF) classifier. Using the extensive CICIoT2023 dataset, which comprises benign traffic and 33 unique attack types, SU was applied to systematically filter out redundant attributes. Empirical evaluations reveal that the integrated SU-RF model, operating on a highly reduced subset of just 15 features (a 68% dimensionality reduction), attains exceptional full-set validation metrics: 99.997% accuracy, 99.996% precision, 99.995% recall, and a 99.995% F1-score. The model's generalization and robustness are further validated through 5-fold and 10-fold cross-validation, which consistently maintained accuracy exceeding 99.35%. When benchmarked against recent state-of-the-art approaches, such as the work by Gheni et al., the proposed framework achieves a 0.55% accuracy improvement while using 53% fewer features, significantly reducing computational costs. Ultimately, this study delivers a highly scalable, computationally lightweight, and robust IDS tailored specifically for the constraints of IoT environments.

Generative Fusion of Two Batik Cual Motifs Using Stable Diffusion

Haris Azhari Ramadhan — 2026-10-06

Batik Cual is a traditional textile heritage from Bangka Belitung that contains distinctive ornamental structures, visual identity, and cultural value. However, previous computational studies on batik motifs have mostly focused on recognition, classification, documentation, or cultural interpretation, while controlled generative fusion for creating new Batik Cual motif variations remains limited. This study proposed a Stable Diffusion-based generative fusion framework to combine two Batik Cual motifs through an image-to-image generation process. The method consisted of motif standardisation, linear image blending as a pre-fusion conditioning image, Stable Diffusion image-to-image generation under fixed parameters, and CLIP-based semantic similarity evaluation. Ten generated motif outputs were produced using the same experimental configuration to ensure a reproducible evaluation process. The results showed that the proposed framework generated varied Batik Cual fusion motifs while preserving several visual characteristics of the input motifs, including dominant reddish colour patterns, ornamental repetition, and decorative structures. The CLIP similarity evaluation produced an average score of 0.2555, with the highest score of approximately 0.3050 and the lowest score of approximately 0.1800. These findings indicate that most generated motifs maintained measurable semantic consistency with the intended Batik Cual fusion concept, although one output showed weaker alignment. Therefore, the proposed framework offers a reproducible and measurable approach for AI-assisted Batik Cual motif exploration.

Power Management Solar-powered Mobile Robot using Interval Type 2 Fuzzy Logic Controller (IT2FLC) Method

Richa Watiasih — 2026-10-06

Mobile robots used for outdoor surveillance require a reliable power source to supply energy to their components, enabling continuous operation without the need for battery replacement. Therefore, an intelligent power management system is essential for mobile robots that utilize solar energy. This study aims to evaluate the performance of solar panels as an energy source for battery charging and power management in such mobile robots. We propose a model of a solar-powered mobile robot system comprising four DC motors, batteries, a buck-boost converter, and solar panels. Battery performance is managed using an Interval Type-2 Fuzzy Logic Controller (IT2FLC). The simulation assesses solar panel output, battery performance, and power management under varying torque conditions for robot movement on flat, uphill, and downhill paths with slope angles of 10°, 20°, and 30°. The simulation results indicate that the maximum power, voltage, and current produced by the solar panel at an irradiance of 1000 W/m² and a temperature of 40°C are 22.71 W, 4.52 V, and 5.02 A, respectively. The State of Charge (SoC) and Depth of Discharge (DoD) are effectively regulated using IT2FLC, with a SoC set point of 50% helping to slow the decrease in battery capacity. The IT2FLC controller's performance strongly influences the solar-powered mobile robot's power management efficiency. Specifically, the robot achieves a power consumption efficiency of 62.02% on flat terrain, 67.82% when traveling uphill on a 10° slope, and 65.78% when moving downhill on a 10° slope.

Spectral Quality-Based Weighted Multi-ROI Fusion for Robust Imaging Photoplethysmography

Salsabila Aurellia — 2026-10-05

Imaging photoplethysmography (iPPG) enables non-contact heart rate (HR) estimation from facial videos, yet its reliability is strongly affected by spatial and temporal signal variability. While many studies focus on improving signal extraction algorithms, the impact of explicit facial region selection and quality-aware fusion remains underexplored. This study systematically proposed a spectral quality-based weighted multi-ROI fusion method and investigates the role of facial region-of-interest (ROI) selection, and temporal quality gating within a controlled and interpretable iPPG framework. A classical pipeline is employed to isolate spatial and quality-related effects. Experiments were conducted on 41 subjects from the UBFC-rPPG dataset. The results demonstrate that quality-weighted multi-ROI fusion provides statistically significant and more consistent performance compared to single-ROI and simple averaging strategies. With moderate temporal gating (τ = 0.20), the proposed method improves estimation accuracy while maintaining high coverage (92.31%) and full subject inclusion. A stricter threshold (τ = 0.25) further reduces the mean absolute error to 2.89 bpm and increases correlation to 0.86, albeit with reduced window-level coverage (41%). These findings highlight a clear trade-off between numerical accuracy and data availability and emphasize the importance of jointly considering spatial reliability and temporal quality control for robust and interpretable iPPG-based heart rate monitoring in realistic real-world deployment scenarios.

A Nonlinear Stochastic Differential Model of EEG Dynamics with Bayesian Change-Point Detection for Early Identification of Brain Abnormalities

Tito Waluyo Purboyo — 2026-10-05

Electroencephalography (EEG) signals exhibit complex nonlinear and stochastic dynamics that fundamentally reflect underlying neural processes. Despite decades of clinical application, conventional EEG analysis remains constrained by linear assumptions or opaque black-box architectures that obscure mechanistic understanding. This study proposes a mathematically grounded nonlinear stochastic differential equation (SDE) framework integrated with Bayesian change-point detection for early, interpretable identification of abnormal brain states. The proposed methodology explicitly characterises nonlinear neural interactions and time-varying stochastic fluctuations through estimation of biophysically meaningful parameters. As a proof-of-concept, validation is performed on data from a single patient (Patient 1, CHB-MIT Scalp EEG Seizure Database) and a single EEG channel (Fp1-F7). Within-patient five-fold cross-validation demonstrates high temporal consistency, with mean sensitivity of 96.8 +/- 0.3% and specificity of 94.2 +/- 0.2%, substantially outperforming conventional linear autoregressive (87.3%, 1.82 s latency) and spectral power density methods (81.5%, 1.35 s). An ablation study confirms that the combined SDE-Bayesian framework yields the highest performance, demonstrating a synergistic gain of +12.4 percentage points in sensitivity over SDE-alone and +8.5 over Bayesian change-point alone. Detection latency of 0.47 s falls within the clinically critical 1-2 second intervention window. Parameter estimation accuracy, validated on simulated data, achieves errors below 2% for all model parameters. Unlike existing approaches, the framework provides physiologically interpretable parameters-neural damping, nonlinearity strength, and stochastic noise level-suitable for mechanistic understanding of pathological brain dynamics. These proof-of-concept results represent a meaningful first step toward clinical deployment; robust generalisation conclusions await future multi-patient and multi-channel validation.

Enhancing Model Trust in Load Forecasting using Counterfactual Explanations

Muhammad Syarif Hidayatullah — 2026-09-02

This study investigates the effectiveness of Counterfactual Explanations (CE) in improving the interpretability of short-term electricity load forecasting models. The forecasting model was developed using Random Forest Regression (RFR) due to its ability to capture nonlinear multivariate patterns in electricity load data and to provide global, descriptive interpretability through feature-importance analysis. However, feature importance provides static and potentially biased explanations, making it insufficient for instance-level interpretability. To address this limitation, the Diverse Counterfactual Explanations (DiCE) framework was integrated to generate multiple “what-if” scenarios that illustrate model-consistent input adjustments capable of shifting forecast outcomes. The study uses hourly electricity load data from Panama collected between 2015 and 2020, with data from 2020 excluded during preprocessing due to atypical demand patterns caused by COVID-19. The model was trained using time-series cross-validation and optimized through grid search. Forecasting performance was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Counterfactual Explanations were assessed using validity, proximity, and compactness metrics, indicating that most generated scenarios are goal-aligned and remain moderately close to the original instances, while the compactness results reflect a feasibility trade-off when large prediction shifts are required. Overall, combining RFR with counterfactual analysis enhances model transparency by bridging global and local interpretability perspectives and provides decision-relevant, scenario-based insights to support transparent and data-driven planning in the energy sector.

Effects of Traffic Direction and IP Type on Dual-Stack Routing Performance in 5G Mesh Networks

Fadli Sirait — 2026-10-25

The coexistence of Internet Protocol version 4 (IPv4) and Internet Protocol version 6 (IPv6) remains important in next-generation communication networks. This study examined the joint effects of traffic direction and IP version on routing performance in a GNS3-based dual-stack mesh emulation designed as a 5G-oriented network-layer abstraction. The model represents multi-hop IPv4/IPv6 routing over high-capacity links and does not emulate the 5G radio access network or 5G Core. A balanced factorial experiment was designed by combining two bandwidth levels, two file sizes, three routing protocols, two traffic directions, and two IP versions. Each of the 48 experimental conditions was scheduled for six repetitions, yielding 288 planned run records. After data verification, 285 valid observations were included in the primary analysis. Throughput, transfer time, end-to-end delay, and packet loss ratio were analyzed using ordinary least-squares factorial models with HC3 robust covariance. Traffic direction significantly affected all four-performance metrics. Download traffic produced higher throughput, shorter transfer time, lower delay, and lower packet loss ratio than upload traffic under both IP versions. IP version significantly affected throughput and transfer time but did not produce significant overall effects on delay or packet loss ratio. Significant Direction × IP Version interactions were observed for throughput, delay, and packet loss ratio, indicating that IPv4–IPv6 performance differences depended on traffic direction. Routing protocol was not statistically significant for any metric; therefore, protocol-level differences were interpreted descriptively rather than as confirmed rankings. These findings demonstrate that dual-stack routing performance is direction-dependent and should be evaluated using traffic-aware factorial analysis rather than separate or aggregated IPv4–IPv6 comparisons.

The Coordinated Energy Management Improvement Microgrid with Hybrid Storage Using Fuzzy Logic

Sri Sukamta — 2026-10-05

The increasing need for electrical energy in residential is caused by the increasing number of houses in rural areas. Currently, the distribution of electrical energy has not reached rural areas, especially in remote areas. Microgrid is a solution in providing electrical energy services for housing. However, the problems of voltage stability, energy management efficiency, energy storage usage period, load fluctuations are the main challenges in the microgrid system. The purpose of this study is to improve the provision of electrical power with an energy management strategy in an energy storage system using fuzzy logic. The method used is to identify load demand, weather data and take a systematic approach by modeling the microgrid system. Furthermore, the proposed energy management system with a coordinated approach, which compares droop control and fuzzy logic controllers. To evaluate the proposed strategy, this study was simulated using MATLAB, based on weather data and load demand. The results of the study showed that the voltage drop and maximum voltage overshot were 2.8 V and 1.6 V respectively in 0.07 S and 0.04 S. Therefore, the proposed strategy shows that there is an increase in efficiency, stability, and reliability when compared to the previous method.

Performance Comparison of Client-Server and Peer-To-Peer Architectures in Local Network Multiplayer Games Based on Unity

Eka Wahyu Hidayat — 2026-10-18

This study aims to compare the performance of Client-Server and Peer-to-Peer Full Mesh architectures in a local network multiplayer game. The research method was conducted through two testing scenarios, namely 2 players and 5 players, with three repetitions of 300 seconds for each session. The parameters measured include latency, throughput, packet loss, jitter, as well as CPU, GPU, and RAM usage. The results show that the Peer-to-Peer Full Mesh architecture has better network performance, with lower latency by 37.4% in the 2-player scenario and 27.7% in the 5-player scenario, as well as higher throughput by 58.2% in the 2-player scenario and 77.6% in the 5-player scenario. The jitter value in this architecture is also lower in both scenarios. Testing of resource usage indicates that the Client-Server architecture is more efficient in CPU usage, with P2P CPU usage being 49.9% higher than Client-Server in the 5-player scenario, while Client-Server uses 23.5% less RAM than Peer-to-Peer Full Mesh. The Client-Server architecture also has lower packet loss by 6.3% in the 5-player scenario. The test results indicate that the Peer-to-Peer Full Mesh architecture is more suitable for scenarios with a smaller number of players that require high responsiveness, while the Client-Server architecture is more appropriate for scenarios with a larger number of players that require resource efficiency and performance stability.

Deep Learning–Based ASD Detection from EEG Signals: A Comparison of InceptionTime and XceptionTime Architectures

Rachmawati Rachmawati — 2026-09-01

Autism spectrum disorder (ASD) is a heterogeneous neurodevelopmental condition characterized by persistent social-communication impairments and restricted or repetitive behaviors. While clinical assessment remains the diagnostic gold standard, the demand for objective and scalable screening tools has motivated EEG-based automated classification. However, under controlled experimental settings, the relative contributions of preprocessing choices and deep learning architecture selection to ASD detection performance remain insufficiently quantified. This study systematically benchmarks artifact-aware preprocessing and model architecture by comparing InceptionTime and XceptionTime across four end-to-end processing schemes that isolate the effects of independent component analysis (ICA) and classifier design under identical segmentation and evaluation protocols. A public King Abdulaziz University EEG dataset comprising 16 subjects (8 ASD, 8 controls) was used. Signals were bandpass-filtered using a fourth-order Butterworth filter (0.5–45 Hz), optionally denoised via ICA, and segmented into 4-s windows with 50% overlap. Models were evaluated using 8-fold subject-wise cross-validation. Performance was assessed using accuracy, precision, sensitivity, specificity, and F1-score, and statistical significance was tested with the Wilcoxon signed-rank test. The Butterworth+ICA+InceptionTime pipeline achieved the best results, with a mean accuracy of 0.9886 ± 0.0046 and an F1-score of 0.9879 ± 0.0049. ICA inclusion and architecture choice yielded significant improvements in accuracy ( for both, ). These findings indicate that structured artifact suppression and multi-scale temporal modeling jointly enhance EEG-based ASD classification, supporting their use in robust clinically oriented screening systems.

Detection of Diseased Salmon Through Advanced Image Processing and Comparative Model Evaluation

Muhammad Abyan Farhan Caesario — 2026-10-13

The rapid spread of diseases in aquaculture poses significant risks to fish health, operational efficiency, and food safety. Manual inspection of fish conditions is time-consuming, subjective, and ineffective for large-scale monitoring. This study presents an advanced image-based detection framework for identifying diseased salmon using modern deep learning architectures, including ResNet-18, EfficientNetV2, Vision Transformer (ViT),

Swin Transformer, and ConvNeXt. The SalmonScan dataset was preprocessed through normalization, resizing, and data augmentation to enhance generalization. Each model was trained and evaluated using accuracy, precision, recall, and F1-score. Additionally, Grad-CAM visualization was applied to interpret model decisions and highlight infection-related regions on fish images. Experimental results demonstrate that transformer-based and next-generation CNN architectures significantly outperform conventional methods, with ConvNeXt and ViT achieving up to 99.59% accuracy. These findings indicate that modern deep learning models combined with explainable visual analysis provide a highly reliable and interpretable solution for automatic disease detection in salmon aquaculture.