Abstract
As power grid construction deepens, communication networks must handle increasingly diverse business traffic. Effective traffic classification is essential for rational resource distribution and service quality. This paper proposes an improved Hidden Naive Bayes (HNB) algorithm for power communication network traffic classification. By leveraging flow-based collection and supervised machine learning, the HNB model addresses the limitations of traditional Naive Bayes while maintaining computational efficiency.
Introduction
The evolution toward "Smart Grids" has transformed power communication networks from simple point-to-multipoint configurations into complex, dynamic systems. These networks carry critical data, including telemetry, control signals, and administrative traffic. Ensuring high service quality and security requires precise knowledge of network traffic types.
Traditional port-based identification is no longer reliable due to dynamic port allocation, and Deep Packet Inspection (DPI) is often too computationally intensive for high-throughput utility environments. Consequently, there is a significant need for a machine learning approach that balances high accuracy with the speed required for real-time power grid management.
Methodology: The Hidden Naive Bayes (HNB) Approach
The standard Naive Bayes classifier operates on the "naive" assumption that all features (attributes) of a data point are independent of one another. In real-world network traffic, this assumption is rarely valid, leading to reduced classification accuracy.
To overcome this, the authors implement a Hidden Naive Bayes (HNB) model. The core innovation involves:
- Hidden Parent Nodes: For every attribute node, a "hidden parent" is introduced. This parent node represents the weighted influence of all other attributes on that specific node, effectively relaxing the independence constraint.
- Conditional Mutual Information: The model calculates weights ($W_{ij}$) using conditional mutual information to determine the degree of influence one attribute has on another. This ensures that significant dependencies are prioritized while filtering out noise.
- Refined Probability Estimation: A variation of m-estimation is used to calculate class probabilities, making the model robust even when dealing with small or incomplete datasets.
Implementation Pipeline
The classification system follows a structured pipeline:
- Data Collection: Traffic is captured at the core router level using Wireshark.
- Flow Construction: Packets are grouped into "flows" based on the standard five-tuple (source/destination IP, source/destination port, and protocol).
- Feature Engineering: Statistical characteristics are extracted from these flows and discretized before being processed by the HNB algorithm.
The study utilized historical traffic for training and real-time captures from an integrated data network's core router for validation, focusing specifically on the unique traffic patterns of electric power supply companies.
Key Insights and Results
The HNB approach offers a "best of both worlds" scenario: the computational speed of Naive Bayes combined with the accuracy of more complex Bayesian Networks. By using conditional mutual information for weighting, the model avoids overfitting while capturing essential feature dependencies.
Simulations comparing the improved HNB algorithm against traditional Naive Bayes revealed:
- Superior Accuracy: The HNB model achieved a higher recognition rate by successfully modeling the interdependence between traffic attributes.
- High Efficiency: The model maintained low computational complexity, satisfying the "timely insight" requirements of power utility network administrators.
- Effective Structure: Results verified that the hidden parent structure effectively relaxes traditional Bayesian constraints, leading to better identification of application types and port services.
Conclusion
The improved Hidden Naive Bayes algorithm provides a robust solution for traffic classification in power communication networks. By addressing feature dependencies through hidden parent nodes, it offers the accuracy needed for modern smart grid management without sacrificing the performance required for real-time operations. Future work may focus on the challenges of encrypted traffic and scaling the model for the broader "Energy Internet."
Summary updated