Deep learning with focal loss approach for attacks classification

Abstract

The rapid development of deep learning improves the detection and classification of attacks on intrusion detection systems. However, the unbalanced data issue increases the complexity of the architecture model. This study proposes a novel deep learning model to overcome the problem of classifying multi-class attacks. The deep learning model consists of two stages. The pre-tuning stage uses automatic feature extraction with a deep autoencoder. The second stage is fine-tuning using deep neural network classifiers with fully connected layers. To reduce imbalanced class data, the feature extraction was implemented using the deep autoencoder and improved focal loss function in the classifier. The model was evaluated using 3 loss functions, including cross-entropy, weighted cross-entropy, and focal losses. The results could correct the class imbalance in deep learning-based classifications. Attack classification was achieved using automatic extraction with the focal loss on the CSE-CIC-IDS2018 dataset is a high-quality classifier with 98.38% precision, 98.27% sensitivity, and 99.82% specificity.