Several works have proposed highly accurate machine learning (ML) techniques for network-based intrusion detection over the past years. However, despite the promising results, proposed schemes must address the high variability of network traffic and need more reliability when facing new network traffic behavior. This paper proposes a new dynamic and reliable network-based intrusion detection model implemented in two phases. First, the behavior of to-be-classified events is assessed through an outlier detection scheme to reject potentially new network traffic, thus, keeping the system reliable as time passes. Second, classification is performed through a dynamic selection of classifier to address the high variability of network traffic. Experiments performed in a new dataset composed of over 60 GB of network traffic have shown that our proposed scheme can improve detection accuracy by up to 33% when compared with traditional approaches.