Abstract
As the number of Internet users and applications increases, application traffic classification is becoming essential for efficient network management. While signature-based and machine learning-based methods exist, Skype—which uses encrypted communication on its own P2P network—remains one of the most difficult applications to identify. This paper proposes a novel method to identify Skype traffic on the fly by examining the packets generated during the login process to build a list of Skype host information {IP, port}. By deploying this system on a campus network, the authors demonstrate high performance and feasibility.
Introduction
Efficient network management requires granular visibility into application-level traffic. However, Skype presents a significant challenge due to its Peer-to-Peer (P2P) architecture, proprietary protocols, and heavy encryption. Traditional Deep Packet Inspection (DPI) fails against encrypted payloads, while Machine Learning (ML) approaches often struggle with real-time performance requirements on high-speed links. This research addresses these gaps by proposing a behavioral identification algorithm that focuses on the unique "fingerprint" of Skype's connection establishment.
Methodology
The proposed system avoids the overhead of decryption by focusing on traffic behavior. It operates using a two-tiered classification logic that dynamically manages a list of active Skype hosts.
Behavioral Identification Algorithm
The core logic involves updating flow information for every incoming packet and checking if the host is already known. If not, the system applies a series of behavioral tests to classify the traffic as Skype or non-Skype.
Login Pattern Analysis
The system's primary strength lies in its ability to identify the Skype login process. Researchers identified a specific 7-step handshake involving Super Nodes (SN), DNS queries, and protocol checks.
By monitoring these sequences, the system extracts {IP, port} pairs and adds them to a dynamic list. Subsequent traffic from these pairs is classified instantly. To maintain efficiency, a "Run Time Action" module removes inactive hosts from the list after 48 hours.
Experimental Evaluation
The system was tested on a Korean campus network over a four-day period, processing approximately 2.5 TB of traffic. To establish ground truth, the researchers used a Traffic Measurement Agent (TMA) installed on end-user devices to verify flow classifications.
Key Results
- Accuracy: The system achieved 100% Precision and 100% Recall within the test environment.
- Latency: Processing time remained under 1 second, even during peak loads.
- Protocol Insights: The study confirmed that 83% of Skype flows used UDP.
- Convergence: The system requires a "warm-up" period of roughly 1.5 hours to populate the host list, after which the false negative rate drops significantly.
Conclusion
This behavioral approach provides a lightweight, high-performance alternative to DPI and ML for identifying encrypted P2P traffic. While the system is highly effective, it relies on the stability of the login protocol. Future work may involve adapting the algorithm to handle protocol evolutions and extending the methodology to other encrypted P2P applications.
Summary updated