Unsupervised algorithm for classifying high BER IP packet based on protocol natural redundancy

This paper addresses transport layer protocol classification for error-prone network environments characterized by bit error rate of 0.1. Without changing existing communication protocol paradigm, we propose robust feature extraction methods and implement a 1D-CNN architecture achieving classification error rates below 10-3. Considering that how to generate labels for training in a noisy background is also an urgent problem, an automated annotation framework leveraging maximum likelihood estimation (MLE) on protocol key fields is developed. Subsequent integration of co-teaching optimization with the 1D-CNN model yields enhanced training performance, attaining 99.8% classification accuracy.