
Intrusion detection system (IDS) research commonly evaluates models on server-grade hardware, overlooking whether these models actually run on the low-cost micro-controllers deployed at the network edge. This study benchmarks five machine learning models–Decision Tree (DT), Random Forest (RF), k-Nearest Neighbors (KNN), Deep Neural Network (DNN), and 1D Convolutional Neural Network (CNN)—for binary intrusion detection across three platforms: ESP32, Arduino Uno, and Arduino Nano. Classical models are exported through MicroMLGen, while neural models are deployed via TensorFlowLite for Microcontrollers with INT8 quantization on ESP32 and through a framework-free float32 PROGMEM approach on 8-bit AVR. Evaluation uses NSL-KDD and Edge-IIoTset v2 under a stratified balanced protocol with multiple random seeds. On NSL-KDD, DT obtains the best trade-off (F1=0.829, 2.65 µs latency on ESP32), outperforming all neural variants. The INT8-quantized DNN deployed on ESP32 shows severe precision-recall imbalance (F1=0.637), whereas the same DNN deployed in float32 on AVR reaches F1=0.762; since host-side INT8 evaluation is lossless and the AVR float32 deployment reproduces host predictions exactly, the on-device gap is a deployment-specific effect of the INT8 TFLite Micro path on ESP32, not of post-training quantization. CNN runs only on ESP32 (F1=0.720). Uniform perfect scores on Edge-IIoTset v2 serve as deployment-context evidence only; the findings above are limited to NSL-KDD. These results challenge the widespread assumption that neural networks automatically surpass classical methods for embedded IDS, and offer practitioners evidence-based guidance.
The rapid growth of online newspapers has increased the effort required to identify and consolidate relevant information across multiple sources. This study proposes an Article News Mining framework that combines TF-IDF-based term extraction, Latent Dirichlet Allocation (LDA), bi-gram phrase enrichment, and objective-driven 5W1H-oriented information extraction. The contribution is framework-level rather than a new topic-modeling algorithm: the pipeline connects thematic discovery, contextual phrase interpretation, and purpose-specific detail retrieval in a transparent workflow. The framework is demonstrated through a retrospective MacBook-related case study and is evaluated by comparing machine-generated summaries with summaries produced by ten respondents across the What, Who, When, Where, and Why dimensions using TF-IDF-weighted text similarity. The preserved research archive contains the aggregate topic and validation outputs, but not the complete article-level collection log or original model-configuration record; therefore, the case study is interpreted as an illustrative proof of concept rather than a fully reproducible benchmark. The reported statistical tests do not identify significant differences for the tested comparisons (p > 0.05), which is interpreted as a failure to detect a difference rather than proof of statistical equivalence. Extracted numerical product claims are separately cross-checked against authoritative product information. The proposed framework provides an interpretable approach for online news mining and decision support while making its validation and archival limitations explicit.
Early detection of breast cancer is essential for reducing mortality, but deploying large deep learning models in resource-constrained clinical settings remains challenging. This study evaluates knowledge distillation (KD) from a ResNet-50 teacher to a MobileNetV3-Small student for classifying breast ultrasound images as benign, malignant, or normal, compared with the same student trained without KD. We partitioned the Breast Ultrasound Images Dataset (BUSI), containing 780 images from 600 patients, using a fixed, stratified 70/15/15 image-level split. Unavailable patient identifiers prevented patient-level splitting, constituting a study limitation. We applied Total Variation Chambolle denoising before training. Grid search selected temperature T = 2 and weighting coefficient α = 0.5, achieving 88.89% validation accuracy. Across five random seeds, the KD student achieved 82.05% ±1.60% test accuracy and 80.93% ±2.04% macro F1-score, compared with 82.74% ±1.64% and 81.88% ±2.17% for the no-KD student, and 83.59% ±4.25% and 82.41% ±5.62% for the teacher. Equivalence testing supported KD–teacher equivalence within a ±5-percentage-point margin (p=0.0498). Paired Wilcoxon tests found no evidence of improvement over the no-KD baseline (p=0.625) or denoising-free training (p=0.4375). The student had 15.5 times fewer parameters and approximately 2.5 times faster inference than the teacher. These findings support compact models for breast ultrasound classification but provide no evidence that KD outperforms simple fine-tuning of the same lightweight architecture on this dataset.
This study proposes an end-to-end framework for evaluating lightweight Convolutional Neural Network (CNN) architectures for maize leaf disease classification in mobile environments. Unlike conventional approaches that primarily emphasize model accuracy, this research integrates performance evaluation with real-world deployment using TensorFlow Lite on Android devices. A dataset of 2,760 maize leaf images across six classes, including a Non-leaf category to improve robustness, was utilized. Five pre-trained CNN models VGG16, ResNet50, MobileNetV2, NasNetMobile, and EfficientNetB0 were evaluated using a consistent experimental setup with 5-fold cross-validation. The results demonstrate that ResNet50 achieves the best performance, reaching a test accuracy of 97.10% and an average K-Fold cross-validation accuracy of 96.38%, confirming its stability and reliability. Furthermore, deployment experiments confirm that the selected model performs real-time inference on Android devices with an average latency of approximately 180 ms and minimal memory overhead, ensuring a seamless, near-instantaneous classification experience for farmers in the field. This study contributes a practical evaluation framework that bridges the gap between deep learning research and real-world implementation in mobile-based plant disease detection.
Software-Defined Networking (SDN) relies on finite switch flow tables, making flow-entry lifetime management relevant to table utilization and controller interaction. This study implements and empirically characterizes a simple, rule-based, three-tier time-out policy in POX, an open-source Python-based SDN controller. A controller-maintained flow-count indicator assigns recorded flow-installation decisions to LOW, MEDIUM, or HIGH, each with predefined idle- and hard-timeout values. The policy was examined in Mininet under nine many-to-one Transmission Control Protocol (TCP) scenarios with 50–500 total hosts, comprising one server and the remaining clients. Reported metrics include cumulative flow-installation events, timeout-category distributions, Packet-In and FlowRemoved counts, controller-derived peak concurrent-flow estimates, and installation events per host. Cumulative flow-installation activity was non-monotonic, reaching its highest observed value of 7,931 events at 350 hosts. HIGH decisions dominated at 50 hosts (65.7%) and reached its highest proportion of 93.7% at 350 hosts. Packet-In activity was also non-monotonic, reaching 10,599 events at 350 hosts, whereas FlowRemoved reached 700 events at 450 hosts. These findings characterize the tested configuration but do not establish optimal capacity or comparative superiority because comparative baselines, repeated trials, controller-resource measurements, independent switch-occupancy validation, and service-level performance measurements were not included.