Machine Learning Metrics for Network Datasets Evaluation

The chapter explores the challenges of using traditional methods like deep packet inspection for network traffic analysis, highlighting the promising results of machine learning in detecting security events even in encrypted traffic. It emphasizes the importance of enhancing ML automation with Active Learning and data drift detection methods to ensure consistent and reliable results across different environments. The authors propose three novel metrics to evaluate the quality and suitability of network traffic datasets, addressing the limitations of existing approaches. These metrics are designed to universally assess linear and non-linear multi-class tasks and can be used to more accurately assess dataset quality over time. The chapter includes a detailed evaluation of the proposed metrics on publicly available datasets, demonstrating their benefits and added value compared to existing solutions. The findings underscore the need for additional dataset evaluation techniques in both scientific research and the production deployment of ML technologies in network security.