CMFTC: Cross Modality Fusion Efficient Multitask Encrypt Traffic Classification in IIoT Environment

Traffic classification, which deduces task-relevant traffic types, is required for network management, security, and quality of service (QoS). In recent studies, deep learning has been used to distinguish encrypted traffic without using handcraft features. However, higher-performance deep learning models often come with a significant computational overhead and parameter count. In this article, we present a novel cross-modal fusion model, CMFTC, which achieves comparable performance to the state-of-the-art (SOTA) model DISTILLER, but with only 18% of its parameters and half the floating-point operations (FLOPs). The CMFTC model is designed to carefully capture the relationship between different modalities and employs a lightweight design for efficient computation. Our model achieves better performance by fusing packet, context, and flow-level modalities. Additionally, we propose a pair of homogeneous student and teacher models with similar architectures, and we adjust the distillation algorithm to improve the student model's inference speed without compromising performance. Our experimental results demonstrate that the proposed model outperforms the current SOTA on several benchmark datasets, including Ton_IoT, IoT-23, ISCX-Tor-2016, ISCX-VPN-2016, and USTC-TFC, with significantly fewer parameters and FLOPs. Furthermore, our model has a better inference speed on real IoT devices, as demonstrated by our experiments.