Enhanced Intrusion Detection in Software-Defined Networks Through Federated Learning and Deep Learning

Abstract

Software-defined networks (SDNs), while offering a revolutionary global view of the network, remain susceptible to a variety of attacks. This vulnerability necessitates innovative solutions for preserving data privacy and enhancing network security. The work presented herein introduces an innovative network anomaly detection methodology leveraging both federated learning (FL) and deep learning (DL) techniques. In contrast to traditional collaborative learning, where potential privacy compromises arise from the distribution of local training data to a central server, the proposed methodology enables each switch in the network to collect data from its connected hosts and independently train a local Long Short-Term Memory (LSTM) model. Subsequently, each switch encrypts and forwards its model parameters to the controller. Upon receipt, the controller decrypts the parameters from each switch, computes their average, and formulates a global LSTM model. This model is disseminated to every switch in the network, enabling each host to retrain its local model according to the global parameters. This iterative process is conducted multiple times to maintain the timeliness of the information. Evaluation of the proposed methodology using the UNSW-NB15 dataset, in conjunction with NF-UQ-NIDSv2 and CICIDS2017 datasets, demonstrated its efficacy in anomaly detection, with performance exceeding a 96.75% accuracy rate.