Software-defined networking (SDN) is a novel approach to network architecture that provides flexibility, dynamism, and adaptability for network management. While accepting these advantages, the primary challenges of this type of network architecture are undoubtedly denial-of-service (DoS) and distributed denial-of-service (DDoS) attacks. For this reason, preventing, detecting, and mitigating these attacks is crucial for the continued operation and efficiency of the network. Studies have shown that statistical approaches, such as entropy, or machine learning methods have attracted the attention of researchers in detecting these attacks. Statistical approaches have low computational overhead and high detection speed, but they have less detection ability than machine learning methods. Conversely, machine learning techniques typically require substantial computational resources and extended convergence times to accurately identify network attacks. In this study, a hybrid approach is proposed that integrates both statistical and machine learning-based methods for the detection of distributed denial-of-service attacks within software-defined networks. This integrated strategy aims to overcome the limitations of each method while capitalizing on their respective advantages. To minimize computational overhead and prevent resource saturation, the architecture employs two dedicated controllers, one responsible for traffic routing and the other for detecting anomalous behavior. Using this architecture, even if the routing controller is disrupted, the attack detection controller can identify the distributed denial-of-service attack. The hybrid model leverages entropy metrics in conjunction with the Packet-in rate over defined time intervals to identify suspicious traffic, which is subsequently subjected to in-depth analysis through machine learning algorithms. Using the proposed approach, the attack detection accuracy rate increases up to 99.47%. Furthermore, due to the examination of a large volume of traffic by statistical methods, which create less overhead, the proposed method shows a reduction in controller computational overhead by up to 10% compared to machine learning methods.