A Federated Learning Autoencoder CNN Hybrid Approach for Privacy Preserving Network Traffic Classification

Federated learning (FL) is an essential technique for classifying network traffic in a decentralized setting while maintaining privacy. In this paper, the Federated Learning Autoencoder Convolutional Neural Network (FL-AECNN), a hybrid federated model, is proposed. For unsupervised feature learning, the FL-AECNN combines a Convolutional Neural Network (CNN) classifier with an autoencoder. This combination improves classification accuracy, generalization, and feature representation quality. Furthermore, it works in environment where the data is Non-Independent and Identically Distributed (non-IID). In the model, the CNN is used for supervised classification, while the autoencoder transforms the traffic features into latent features. This model is hybrid due to its two functionalities. For the experiment, a customized Android traffic dataset with ten classes and the ISSC VPN2016 benchmark dataset with five classes are used to test the proposed model. In the experiment, the SMOTE algorithm is employed to balance the classes, and log transformation is used to normalize all the datasets to address the skewed features. The Federated Averaging algorithm, or FedAvg, is used to aggregate the model globally, while the local model is trained independently by the ten clients and the central server. The average training accuracy of the FL-AECNN model is 90.24%, and the range of the testing accuracy of the model is between 77.01% and 82.54%. These results show that the FL-AECNN model performs better in terms of accuracy and consistency compared to the Federated Learning Convolutional Neural Network (FLCNN). These results indicate the possibility of applying federated supervised classification and federated unsupervised representation learning to develop a new method of safe traffic assessment.