With recent advancements in artificial intelligence (AI) and next-generation communication technologies, the demand for Internet-based applications and intelligent digital services is increasing, leading to a significant rise in cyber-attacks such as Distributed Denial-of-Service (DDoS). AI-based DoS detection systems promise adequate identification accuracy with lower false alarms, significantly associated with the data quality used to train the model. Several works have been proposed earlier to select optimum feature subsets for better model generalization and faster learning. However, there is a lack of investigation in the existing literature to identify a common optimum feature set for three main AI methods: machine learning, deep learning, and unsupervised learning. The current works are compromised either with the variation of the feature selection (FS) method or limited to one type of AI model for performance evaluation. Therefore, in this study, we extensively investigated and evaluated the performance of 15 individual FS methods from three major categories: filter-based, wrapper-based, and embedded, and one ensemble feature selection (EnFS) technique. Furthermore, the individual feature subset’s quality is evaluated using supervised and unsupervised learning methods for extracting a common best-performing feature subset. According to our experiment, the EnFS method outperforms individual FS and provides a universal best feature set for all kinds of AI models.