Edge network controllers must allocate bandwidth under rapidly changing traffic demand while avoiding both underprovisioning and excessive overprovisioning. This paper presents a forecast-driven adaptive bandwidth allocation frame-work that converts short-horizon traffic predictions into edge resource-control decisions evaluated in an analytical QoS simulator. Using a PCAP-derived traffic corpus, reactive rolling-mean allocation, persistence-based allocation, LSTM-driven allocation, CNN–LSTM-driven allocation, a TensorFlow Lite-compatible CNN–LSTM policy, hybrid max policies, adaptive safety-factor controllers, and a perfect-forecast + SF reference are compared. Results show that forecast-driven allocation reduces average latency, packet loss, jitter, and SLA violations compared with reactive control when forecast bias is favorable, while adaptive safety-factor tuning reduces underprovisioning without unbounded bandwidth waste. Persistence and standalone LSTM fixed-safety policies are reported separately as burst-sensitive failure modes; they are excluded from primary charts and capped-comparison tables because rare underprediction produces extreme uncapped analytical delays. The analysis further shows that allocation quality depends not only on RMSE but also on forecast bias, underprediction rate, and safety-factor behavior. This study does not introduce a new forecasting architecture; forecasting models are used as input predictors. The simulator is analytical and intended for comparative policy evaluation, not live deployment measurements.
Emannuel T. Saligue, Patrick D. Cerna· 2026 International Conferenc...· 0 citations
The increasing complexity of modern network environments has heightened the need for practical and scalable intrusion detection solutions. This study presents the development of a real-time, flow-based Intrusion Detection and Security Management System (IDSMS) that integrates live traffic monitoring, machine learning-based intrusion analysis, centralized data storage, and web-based visualization. The system captures network traffic from a MikroTik-based gateway using a passive flow sniffer deployed on a Kali Linux monitoring node. Packets are aggregated into bidirectional flows and transformed into statistical feature vectors, which are analyzed by a machine learning inference service through an application programming interface. Detection results, including risk level and attack classification, are stored in a centralized database and presented through an interactive dashboard. The intrusion detection engine employs a two-stage ensemble approach: Stage A performs binary classification (benign vs. malicious) using a calibrated stacking ensemble of Random Forest, Gradient Boosting, and XGBoost classifiers with Logistic Regression as meta-classifier, achieving 99.92% accuracy and 0.99987 ROC-AUC. Stage B performs multi-class classification across 14 attack categories using a Random Forest classifier, achieving 99.83% accuracy with per-class F1-scores exceeding 99.6%. Development-stage testing using simulated attack traffic demonstrates the system’s ability to capture, analyze, and visualize network threats in real time, with average inference latency of 12.4 ms per flow and processing throughput of 3,842 flows per minute. The proposed system provides a practical foundation for deployment in shared-access networks, small enterprises, and IoT-enabled environments.
Jimson A. Olaybar, Patrick D. Cerna· 2026 International Conferenc...· 0 citations