Reducing Mean Absolute Error in Attack Predictions with Power-Parametrized Choquet Integrals
Denner G. Ayres, Abreu Quevedo, Graçaliz Pereira Dimuro, Giancarlo Lucca, Bruno L. Dalmazo · 2025
Computer networks are essential to ensure fast and reliable access to information, both in business and everyday life. With the growing number of connected devices, it has become increasingly important to monitor and manage data traffic effectively. This work explores methods to predict network traffic and detect unusual behavior that could indicate cyberattacks. We tested moving average models, which are known for having low dependency on historical data. The model with the lowest prediction error was then compared to an aggregation function based on the Choquet integral, adjusted using different mathematical measures. The results showed that adjusting the size of the sliding window can improve prediction accuracy. Functions like entropy and relative amplitude performed especially well with larger window sizes. The proposed approach does not require prior training, making it suitable for real-time use in dynamic network environments. The flexibility of this combined approach demonstrates potential for practical applications in network security. In different evaluated scenarios, the method achieves a precision of 99% and, in others, an accuracy of 98%.