Traffic Refinery

Network management often relies on machine learning to make predictions about performance and security from network traffic. Often, the representation of traffic is as important choice model. The features that model on, those features, ultimately determine accuracy, well where whether can be deployed in practice. Thus, design evaluation these models requires understanding not only accuracy but also systems costs associated with deploying an operational network. Towards this goal, paper develops a new framework system enables joint both conventional notions (e.g., accuracy) systems-level different representations We highlight two dimensions for practical tasks, video streaming quality inference malware detection, demonstrate importance exploring find appropriate operating point. benefit range present Traffic Refinery, proof-of-concept implementation monitors at 10~Gbps transforms real time produce variety feature learning. Refinery highlights space makes it possible explore learning, balancing related extraction training against accuracy.