Reproduce “CATO: End-to-End Optimization of ML-Based Traffic Analysis Pipelines” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Gerry Wan, Shinan Liu, Francesco Bronzino, Nick Feamster, Zakir Durumeric

CATO, an end-to-end framework, systematically optimizes both predictive performance and system costs for ML-based network traffic analysis pipelines by combining multi-objective Bayesian optimization with direct real-world measurement. It identified Pareto-optimal solutions that achieved up to 3600 times lower inference latency and 3.7 times higher throughput compared to traditional baselines, while maintaining or improving model accuracy across various traffic analysis tasks.

Key insights

  • Jointly optimizing both predictive performance and system costs through a multi-objective approach is essential for developing deployable ML-based network traffic analysis solutions.
  • Direct, end-to-end measurement of pipeline costs and performance, rather than relying on heuristic approximations, is critical for accurately guiding the optimization process and validating practical deployability.
  • Bayesian Optimization, enhanced with techniques like dimensionality reduction and prior injection, can efficiently explore complex, high-dimensional search spaces for optimal feature sets and connection depths in network ML pipelines.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.