Network Log Implementation for GRU Based Bandwidth Classification

Network bandwidth management using log data is a challenging task, especially in anomaly detection, e.g., fraudulent bandwidth that violates the Service Level Agreement (SLA). The present study suggests a deep learning automatic classification method for network logs, which leverages the Gated Recurrent Unit (GRU) and is used in time-series tensor configurations given as [N, 5, 15]. Data was gathered in real time during 29 days with the aid of a MikroTik RB1100AHx router, and it created more than 867,000 rows of data with three logs per second. The logs were classified into three classes: Genuine, Fake, and No Heavy Activity. Pre-processing involved windowing sequences, normalisation, and SMOTE balancing, whereas the GRU model comprised update and reset gates, followed by a Dense layer and a Softmax 3-class output. The model was trained with categorical cross-entropy loss and optimized with the Adam optimizer, validated with a 5-fold cross-validation strategy. The results achieved a 86.8% mean accuracy and an F1 score of 0.90 in the classification of Genuine Bandwidth, indicating that the GRU can successfully detect temporal patterns in network logs. This system is locally deployable through the G-Radio interface, demonstrating its feasibility, scalability, and substantial contribution to automatic bandwidth classification without packet inspection.