Title: Neural network optimisation combining feature filtering and cross entropy in software defined network security

Authors: Lu Liu

Addresses: School of Information Engineering, Shaanxi Institute of International Trade and Commerce, Xi'an, 712046, China

Abstract: Software defined networks (SDN) are an emerging network architecture with high flexibility and editable capabilities. However, the centralised control plane of SDN makes it vulnerable to abnormal traffic attacks, while traditional detection methods face challenges such as feature redundancy and data imbalance. To improve the stability and security of SDN, this study proposes a lightweight federated learning-based SDN anomaly detection model that combines a feature filtering module with a cross-entropy loss function optimisation. The results showed that after five iterations, the loss values of all three models reached convergence. The federated learning model without compression had the worst convergence effect, and the convergence of the two models trained 20 and 15 times was basically the same. After completing the model training, the loss values of these three models remained around 1.0. The software defined network abnormal traffic detection model could reduce the loss value to around 1.0 during training, maintain recall and accuracy at around 0.99, and maintain precision at around 0.98. The software defined network abnormal traffic detection model can effectively identify attack behaviours in the network, improve the security protection level, and protect the privacy of users during network use.

Keywords: software defined network; SDN; deep learning; cross entropy; feature selection; abnormal traffic.

DOI: 10.1504/IJCC.2025.151117

International Journal of Cloud Computing, 2025 Vol.14 No.4, pp.353 - 370

Received: 09 Apr 2025
Accepted: 11 Jun 2025
Published online: 14 Jan 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article