A Framework for Elephant Flow Detection for SDNs Based on the Classification


ARABIAN JOURNAL FOR SCIENCE AND ENGINEERING, no.3, pp.4243-4252, 2024 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Publication Date: 2024
  • Doi Number: 10.1007/s13369-023-08345
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Aerospace Database, Communication Abstracts, Metadex, Pollution Abstracts, zbMATH, Civil Engineering Abstracts
  • Page Numbers: pp.4243-4252
  • Karadeniz Technical University Affiliated: Yes


Recently, there has been considerable interest in the classification of flows among existing applications in SDN, such as traffic engineering, service quality, and network management. Particularly, the classification methods used for elephant flow detection play an essential role. With elephant flow detection, the high bandwidth provided by data centers is used effectively, and traffic routing can be performed on the network without delay. The most important goals in elephant flow classification are to detect elephant flows in the shortest possible time and with high accuracy. This study aims to lead further studies by analyzing different classifiers for elephant flow detection, which is the preliminary step of elephant flow routing in SDN. For this purpose, a new framework is proposed for elephant flow detection. The framework is tested with six different classifiers; most have not been used in the literature before in elephant flow detection on SDN. The proposed framework's first step is to convert UNI1 and UNI2 datasets containing actual network data into flowing datasets using the Flowrecorder tool. The OTSU thresholding method is used for ground-truth labeling. Then, feature extraction is made for the six different classifiers. In the proposed framework, 11 various features are extracted for each flow. In addition, the most suitable parameters for each classifier are determined by the grid search method. The success of classification techniques is calculated using accuracy, precision, recall, F-score, and running time measures. When the results are examined, the proposed framework obtained pretty successful results in SVM and decision tree methods.