Model Uncertainty Based Unsupervised Real-Time Drift Detection in Network Traffic Classification
Minyao Liu, Pan Wang, Yingchun Ye, Xuejiao Chen · 2024
In recent years, deep learning-driven techniques for traffic classification have shown superior efficacy. However, the volatility of network environments means that the potential drift in data distribution (DDD) poses a significant threat to the reliability of these classification systems. Despite the proliferation of research on methods to detect such drifts, the majority necessitate labels for real-time flow data. Moreover, some unsupervised methods result in high resource consumption and latency when applied to network traffic. To address these challenges, we introduce an unsupervised real-time drift detection method based on model uncertainty (MU-UDD). This approach uses a Dirichlet distribution to represent confidence levels associated with outputs from the classifier. It computes the entropy of this distribution to gauge model uncertainty, thereby providing a reliable indicator for the presence of drift. We employ a sliding window mechanism coupled with a weighted CUSUM-style method to track entropy fluctuations, thereby swiftly detecting DDD occurrences. Experiments on different datasets have validated the superiority of MU-UDD over competing methodologies in reducing false positives and negatives while enhancing precision in drift location identification.