Data Augmentation for Network Security

Défense de mémoire d'Antoine Petit

 

In the last decade, with the emergence of new technologies, the industry 4.0, the spread of IoT, ... Network security becomes more and more challenging. Network anomalies detection gain in interest but network datasets are unbalanced where the anomalies are lost in normal samples. The Classification of unbalanced datasets tends to deteriorate the score of the minority classes. A well-known way to counter this unbalance problem is the data augmentation. Well known in the images domain, less in the network, this study explores the data augmentation for network security datasets. Two techniques of augmentation are compared, a variant of the Generative Adversarial Network (GAN) as the literature's shown that a naive adaption of GAN does not work well in the networks domain, Divide Augment Combine (DAC) with GAN is explored in this study. The second technique is the data transformation, what transformations can be applied on network samples without breaking the semantic. These two techniques were trained and tested on the UNSW-NB15 dataset.

 

Keywords: anomalies detection, GAN, data generation, data augmentation, data transformation.

Contact : Isabelle Daelman - isabelle.daelman@unamur.be
Télecharger : vCal