Leveraging Federated Learning and XAI for Privacy-Aware and Lightweight Edge Training in Network Traffic Classification

The exponential growth of internet traffic causes significant challenges for network traffic classification, such as maintaining data privacy and requiring more computing resources. To address such challenges, moving the processing to data instead of the opposite shows some promise. Related technologies, such as edge computing, present a potential solution as they reduce latency and bandwidth and consider data privacy. However, edge computing often runs on resource-constraint devices, thus making complex model training like deep learning challenging. As such, this paper proposes a distributed and lightweight edge training method for network traffic classification using federated learning and XAI, which are currently underexplored in the domain. In doing so, we managed to run deep learning on edge devices and preserve data privacy. We evaluate the proposed method using the ISCXVPN2016 dataset on three NVIDIA Jetson devices to emulate edge devices. The results show a 36.71% improvement in F1 score, training time, memory, and GPU utilization when benchmarked against the more conventional centralized training.