Encrypted Traffic Classification Based on an Improved Clustering Algorithm

Classification analysis of network traffic based on port number or payload is becoming increasingly difficult from security to quality of service measurements, because of using dynamic port numbers, masquerading and various cryptographic techniques to avoid detection. Research tends to analyze flow statistical features with machine learning techniques. Clustering approaches do not require complex training procedure and large memory cost. However, the performance of clustering algorithm like k-Means still have own disadvantages. We propose a novel approach of considering harmonic mean as distance matric, and evaluate it in terms of three metrics on real-world encrypted traffic. The result shows the classification has better performance compared with the previously.