Data Drift Detection in Network Data

Deep learning models have significantly advanced various fields by autonomously recognising complex patterns within large datasets, enabling tasks previously considered unachievable. Despite their strengths, these models face critical challenges when deployed in dynamic real-world environments due to their reliance on static, historical data. Real-world conditions are constantly evolving and causing shifts in data distribution, a phenomenon known as data drift, which negatively impacts model performance and reliability. This thesis specifically addresses data drift within network traffic classification. It systematically investigates and evaluates state-of-the-art techniques for detecting out-of-distribution (OOD) network data and effectively managing shifts in data distribution. Through extensive experiments using a robust 1D-CNN model on real network datasets, this research identifies effective methods for detecting and mitigating data drift. Furthermore, an interactive human-in-the-loop dashboard is developed that enables domain experts to visually detect, analyse and manage data drift. This tool facilitates expert-driven decision making and supports continuous model refinement. This work significantly advances the theoretical and practical understanding of data drift, providing valuable insights and robust solutions for dealing with evolving data distributions. As a result, the work contributes to the development of resilient and adaptive deep learning systems capable of maintaining high accuracy and reliability in dynamic network environments.