<![CDATA[Attackers use malware to launch attacks in the internet and corporate networks. Over the years, machine learning techniques have been found promising for the classification of these attacks because they have the ability to identify unknown threats. Botnets are networks of compromised devices and have been found to be powerful threat vectors that are used against modern systems because they use command and control (C2) characteristics which make their detection very difficult. Generally, to build attack detection models, intrusion datasets are employed. Comprehensive study of the benchmarking datasets used in intrusion detection researches can provide different actionable insights to other researchers. There have been studies that investigated the analyses of datasets for building intrusion detection systems. However, there has been less focus on the analysis of intrusion detection datasets that are used specifically for botnets detection. This study reported an overview of a popular botnet dataset named CTU-13. Thereafter, the work carried out detailed exploratory analysis of the dataset. The study equally sought to identify if the dataset is representative enough for Machine Learning based botnet detection studies. All the thirteen scenarios in the dataset were used for the experimentations. The exploratory analyses were carried out on each of the thirteen scenarios of the dataset with a view to gaining better understanding of the patterns and characteristics of data in each of them. The information obtained from the overview and exploratory analyses provided actionable insights on how to better use the datasets for improved botnet classification. The challenges of using the captures of the dataset were also identified. In particular, the exploratory investigation of the thirteen captures of the CTU-13 dataset revealed that it has very complex patterns, contain mixed data types and suffers from high class imbalance problem. The results of the exploratory analyses can guide the decision of future cyber security researches. Thus, improved machine learning-based botnet detection models can be built by attending to the issues in the dataset.]]>
