Expert Selection for Wordlist-Based DGA Detection

Domain Generation Algorithms (DGAs) have evolved beyond traditional pseudorandom patterns, with wordlist-based variants generating linguistically coherent domains that evade conventional detection methods. While previous research has primarily focused on generalist detection approaches across multiple DGA types, systematic expert model selection specifically targeting wordlist-based variants remains largely unexplored. This work addresses expert model selection for wordlist-based DGA detection, where expert models refer to specialized architectures trained exclusively on specific DGA categories. We conduct systematic evaluation of seven candidate models across transformer, convolutional neural network (CNN), and traditional machine learning approaches. Models were trained on a balanced dataset of 160,000 domains spanning eight wordlist-based DGA families and evaluated using a rigorous two-phase protocol that measures both performance on training families and generalization to previously unseen variants. Our comparative analysis identifies fine-tuned ModernBERT as the optimal expert model, achieving 86.7% F1-score on known families while maintaining 80.9% performance on unknown families with 26ms inference time on NVIDIA Tesla T4 GPUs, enabling processing of approximately 38 domains per second. The study validates that domain-specific expert training significantly outperforms generalist approaches trained on diverse DGA families, with F1-score improvements of 9.4% on familiar variants and 30.2% on unseen families. This performance gain indicates that focused expertise develops transferable linguistic patterns rather than memorization of specific family characteristics.