The number of Internet of Things (IoT) devices is growing quickly, creating large networks that are attractive targets for malware such as Mirai and Mozi. Most existing epidemic models assume homogeneous mixing, where every device has an equal probability of contacting others, although real IoT networks are characterized by heterogeneous connectivity in which highly connected hub devices accelerate malware propagation. This study proposes a network-based SEIRS model that incorporates connectivity-driven transmission dynamics by allowing the infection rate to depend on device degree. The model is used to derive the basic reproduction number (R0) and establish the local and global stability conditions of the malware-free and endemic equilibria. Three intervention measures, namely patch deployment, network quarantine, and user-awareness campaigns, are incorporated into an optimal control framework using Pontryagin's Minimum Principle to balance infection reduction and intervention costs. The proposed model is evaluated using CICIoT2023 and CIC IoT-DIAD 2024 traffic datasets, complemented by network-topology information from N-BaIoT, representing 105 IoT devices and 33 attack types. Simulation results show that the combined optimal-control strategy reduces the peak number of infected devices by 78.4% compared with the uncontrolled scenario and by 31.6% compared with a conventional SEIRS model that ignores network structure, while reducing overall intervention costs by 42.3%. Sensitivity analysis identifies device connectivity as the dominant factor influencing outbreak severity, highlighting the importance of network-aware segmentation and prioritized patching for effective IoT malware containment.
This study evaluated the initial feasibility and practicality of Educloud, a cloud-based digital learning platform integrating Learning Management System functions, digital assessment, learner-progress monitoring, and an AI-supported Edupreneurship workflow. Educloud was developed using Laravel and PostgreSQL, with AI functionality provided through the Google Gemini Developer API using the gemini-2.5-flash model. The AI component processes pseudonymised learner and team profiles to generate business-idea recommendations and formative feedback. The study employed a research and development approach based on an adapted ADDIE model, involving three expert validators and ten users in a small-group trial. Functional testing assessed API connectivity, contextual-data processing, privacy protection, response validation, storage, and error handling. Expert validation produced a score of 489 out of 525 (93.14%), a mean of 4.66, and an Aiken’s V of 0.91. The small-group trial produced a score of 1,117 out of 1,250 (89.36%), a mean of 4.47, and a Cronbach’s alpha of 0.93. The findings indicate that Educloud demonstrates strong initial feasibility and practicality. Further field testing is required to evaluate the accuracy of AI-generated recommendations and their effects on learning and Edupreneurship outcomes.
The rapid expansion of digital multimedia distribution has increased the demand for dependable copyright protection and content authentication. Digital image watermarking has emerged as a principal solution, yet the fundamental challenge remains in balancing imperceptibility and robustness against various signal processing and geometric attacks. This paper proposes a novel hybrid watermarking scheme that combines the Integer Wavelet Transform (IWT), the Discrete Cosine Transform (DCT), and Singular Value Decomposition (SVD), while entropy and edge entropy decide where the watermark is embedded. The use of IWT overcomes the floating-point precision loss inherent in conventional Discrete Wavelet Transform (DWT), ensuring perfect reconstruction and higher accuracy in watermark extraction. DCT is employed to leverage energy compaction in mid frequency bands, while SVD provides exceptional stability against geometric distortions due to the inherent invariance of singular values. The critical innovation lies in the integration of entropy and edge entropy metrics to dynamically identify highly textured image blocks with maximum natural masking capacity. Unlike conventional HVS or variance-based methods that often mischaracterize directional edges, the proposed dual-entropy model precisely quantifies both luminance obliviousness and structural sensitivity, enabling the embedding of stronger watermarks without compromising perceptual quality. Extensive experiments on eight standard 512×512 benchmark images demonstrate consistently high imperceptibility, with an arithmetic average PSNR of 40.46 dB, SSIM of 0.9965, and FSIM of 0.9995. Robustness evaluations under fourteen categories of signal processing and geometric attacks yield Normalized Correlation (NC) values predominantly above 0.99, outperforming four recent state-of-the-art methods in eleven of twenty attack scenarios. The identified limitations include reduced robustness against large-angle rotation attacks, an inherent SVD challenge. The proposed framework demonstrates significant potential for secure, imperceptible digital image watermarking.
Pulmonary tuberculosis (TB) remains a major public health challenge, especially in developing countries where early and accurate diagnosis is critical. Although deep learning models have shown promising performance in TB detection using chest X-ray (CXR) images, most studies rely on large-scale public datasets that may not adequately represent the variability and limitations of real-world clinical data. This study aimed to evaluate the robustness and generalization capabilities of the CNN-VGG19 architecture when applied to small-scale clinical datasets. The dataset consisted of 434 CXR images collected from Haji Makassar Hospital, Indonesia, and categorized into tuberculosis, pleural effusion, and normal classes. A transfer learning approach was applied, and model performance was evaluated using analyses of accuracy, precision, recall, F1 score, and confusion matrices, with comparisons to ResNet50 and InceptionV3 under identical experimental conditions. The results showed that CNN-VGG19 achieved an accuracy of 0.8807 and a macro F1 score of 0.8800, outperforming the comparative models. These findings suggest that CNN-VGG19 provides stable performance in a data-constrained environment and highlight the impact of class overlap on classification performance.
Chili cultivation is considerably affected by leaf diseases, making timely and reliable disease identification essential for maintaining agricultural productivity. While deep learning has demonstrated strong capabilities in plant disease recognition, many existing studies primarily focus on modifying network architectures rather than optimizing the training process. This study introduces a two-stage transfer learning framework that combines frozen-backbone adaptation with controlled fine-tuning to enhance classification performance. The experiments utilized a publicly available dataset comprising 12,000 augmented chili leaf images representing six disease categories. Several deep learning architectures were investigated, including DenseNet121, EfficientNetB0, MobileNetV2, ResNet152V2, and the Vision Transformer (ViT-B16). The experimental results indicate that the proposed two-stage training strategy improves classification performance consistently across the evaluated architectures. Among the models, ViT-B16 obtained the highest test accuracy of 99.61%, together with stable macro-averaged precision, recall, and F1-score values. Furthermore, the confusion matrix and ROC analysis revealed strong discrimination among the disease classes, with only limited errors occurring between visually similar categories. These findings highlight the importance of optimizing the training schedule when adapting pretrained models for plant disease recognition. Overall, the proposed framework provides an effective and scalable approach for chili leaf disease classification while offering an alternative to approaches that primarily rely on architectural modifications.
