The article presents the development of an integrated mathematical model for forecasting network load and detecting cybersecurity anomalies, based on modern deep learning methods and the autoencoder architecture. The proposed approach combines neural network-based forecasting functionality with automated mechanisms for identifying deviations in network traffic behavior. At the initial stage, the model performs preprocessing of historical data using normalization and exponential smoothing, which allows for the effective extraction of current load patterns. Forecasting is carried out using a deep neural network optimized by gradient descent to minimize the mean squared error (MSE). An autoencoder is applied for anomaly detection, trained on normal data and employing the Euclidean norm of the difference between input and reconstructed signals to quantify the anomaly level. Anomaly boundaries are adaptively generated based on standard deviation and a sensitivity parameter, enhancing detection accuracy under dynamic network conditions. Additionally, the study introduces a model for assessing the criticality of network states by considering the proportion of anomalous values within the o
📖 افتح في inklap 🔗 DOI 📮 اطلب بحثاً