CATO

Overview

CATO jointly optimizes both the predictive performance and systems costs of ML-based traffic analysis serving pipelines.

The Problem

ML-based network analysis solutions often only optimize predictive performance, overlooking practical challenges of running models against network traffic in real-time. Efficiency of the serving pipeline is critical for usability.

How CATO Works

Uses multi-objective Bayesian optimization to efficiently identify Pareto-optimal configurations. Automatically compiles end-to-end optimized serving pipelines deployable in real networks.

CATO searches for optimal combinations of:

• Which network traffic features to analyze
• How much traffic data to collect before predictions

Results

Compared to popular feature optimization techniques:

• Up to 3600× lower inference latency (minutes to <0.1 seconds)
• 3.7× higher zero-loss throughput
• Better model performance simultaneously

Example: IoT device classification achieved better accuracy using just 3 packets vs. baseline approaches requiring 10 packets.

Resources

• Blog Post

Citation

@article{wan2025cato,
  title={CATO: End-to-end Optimization of ML Traffic Analysis Pipelines},
  author={Wan, Gerry and Liu, Shinan and Bronzino, Francesco and Feamster, Nick and Durumeric, Zakir},
  journal={USENIX Symposium on Networked Systems Design and Implementation},
  year={2025}
}