Predicting Cyberattack Duration in Next Generation Networks: A Novel Transformer-Based Approach

In the face of increasingly complex cyberattacks, particularly within 5G networks, accurately predicting the duration of an ongoing attack has become essential for effective attack mitigation. In this work, we tackle this issue by exploring the use of deep learning models to forecast cyber attack duration, thus enabling improved resource allocation and mitigation strategies. Using the diverse UNSW-NB15 dataset, our approach proposes data transformation and the creation of new key features, modeling the problem as a time-series forecasting one to predict the remaining attack time of ongoing attacks. Subsequently, several deep-learning models, suited for time-series data, have been designed. Based on automated hyperparameter tuning and feature engineering, we further refine the developed models. Through extensive simulations, we evaluate the performance of developed approaches in terms of mean absolute error (MAE). Obtained results indicate that the Transformer-based model outperforms other methods by achieving the lowest validation MAE with nearly 60 % reduction compared to the well-known Long Short-Term Memory (LSTM) model, thus showcasing its robustness in accurately predicting the duration of any type of attacks.