CATO - End-to-End Optimization of ML-Based Traffic Analysis Pipelines

Learn about CATO, a groundbreaking framework that addresses the practical deployment challenges of machine learning-based network traffic analysis systems in this 12-minute conference presentation from NSDI '25. Discover how researchers from Stanford University, University of Chicago, and ENS Lyon tackle the critical gap between ML model performance and real-world deployment efficiency in network traffic analysis applications. Explore the limitations of current ML approaches that focus solely on predictive performance while neglecting the systems costs and efficiency requirements of real-time network traffic processing. Understand how CATO revolutionizes this field by jointly optimizing both predictive performance and serving pipeline efficiency through multi-objective Bayesian optimization techniques. Examine the framework's ability to automatically identify Pareto-optimal configurations and compile end-to-end optimized serving pipelines ready for deployment in production networks. Analyze the impressive performance improvements demonstrated in evaluations, including up to 3600× reduction in inference latency and 3.7× increase in zero-loss throughput compared to traditional feature optimization techniques, all while maintaining or improving model accuracy. Gain insights into how this research bridges the gap between academic ML research and practical network operations, making ML-based traffic analysis solutions more viable for real-world deployment.