Extended Abstract: A Transfer Learning-Based Training Approach for DGA Classification

We propose a novel training method based on transfer learning (TL) and fine-tuning to improve the detection performance of domain generation algorithm (DGA) classifiers. This new training approach allows a classifier to learn nuances of domains generated by specific DGAs, which greatly improves a classifier’s separation performance between benign and malicious samples. Additionally, we develop and optimize four novel models for DGA binary and multiclass classification based on more recently proposed deep learning (DL) architectures. We comparatively evaluate the resulting DGA classifiers in a unified setting for statistically significant improvements, and assess whether the classifiers generalize well between different networks and are time-robust. In a real-world setting, our best performing model improves the state of the art by 2–3.4% in true-positive rate (TPR) at the same fixed false-positive rates (FPRs).