Forecasting anomalies in network traffic

The increasing volume of traffic, growing number of connections in telecommunication networks, and rising number of mobile devices place significant demands on network providers. These challenges can lead to congestion, latency issues, and security vulnerabilities. However, they can be mitigated or even prevented by identifying network failures in advance. Anomaly detection plays a crucial role in proactively addressing these issues, enabling network operators to optimize network performance, enhance security, and improve the overall user experience. In this paper, a method for predicting anomalies based on machine learning has been implemented. The LSTM-SMOTE model, which was trained and tested on the KDD-NLS dataset, was considered and the results of the forecasting model were analyzed. Developing the multi-classification model proved to be a challenging task, primarily due to the limited number of attack types. SMOTE is designed to address such difficulties. Imbalanced datasets present a major challenge in predictive modeling, especially when solving classification problems. The four main types of attacks include Denial of service (DoS) attacks, Probe attacks, Privilege attacks, and Access attacks. In this work, three neural network models were developed, including: binary classification, four-class classification, multi-class classification. It is observed that the prediction model retrained again showed the best results, then the model trained with new anomalous data. The LSTM-SMOTE multiclass model achieved the highest performance, with its predictive accuracy rising from 75 % to 99 % across iterations, underscoring its strong dependence on the quality and quantity of data. Practical application of the results obtained can be applied for optimizing network performance