Today unauthorized access to sensitive information and cybercrimes is rising because of increasing access to the Internet. Improvement in software and hardware technologies have made it possible to detect some attacks and anomalies effectively. In recent years, many researchers have considered flow-based approaches through machine learning algorithms and techniques to reveal anomalies. But, they have some serious defects. By way of illustration, they require a tremendous amount of data across a network to train and model network’s behaviors. This problem has been caused these methods to suffer from desirable performance in the learning phase. In this paper, a technique to disclose intrusions by Support Vector Regression (SVR) is suggested and assessed over a standard dataset. The main intension of this technique is pruning the remarkable portion of the dataset through mathematics concepts. Firstly, the input dataset is modeled as a Directed Graph (DG), then some well-known features are extracted in which these ones represent the nature of the dataset. Afterward, they are utilized to feed our model in the learning phase. The results indicate the satisfactory performance of the proposed technique in the learning phase and accuracy over the other ones.