Lightweight Multi-Input Shape CNN-based Application Traffic Classification

This research focuses on the input shape of CNN-based application traffic. The previously proposed multi-input model CNN classification method classified applications through various shapes of features derived from fixed-length packets, achieving a higher classification accuracy compared to traditional CNNs. However, it had limitations such as vulnerability to overfitting despite its high classification accuracy and slow inference speed. To overcome these challenges, we introduce a lightweight version of the previously proposed MISCNN, called MISCNN+. MISCNN+ demonstrated approximately 2.9 times faster inference speed and a 3.6% improvement in classification accuracy compared to the previous version.