Accuracy is only one part of a deployable classifier
A research note from the Adaptive AI Lab — production network classifiers must trade off update time, memory, latency, privacy, and behavior on unknown traffic, not only benchmark accuracy.
Research note · August 2026 · Production ML
A model that tops a benchmark can still fail in a live network. Real deployments also care about update time when new applications appear, memory on middleboxes, latency per flow, privacy of the underlying traffic, and behavior on traffic the model has never seen.
Our work measures those trade-offs explicitly, so a model can be chosen for the operating environment — an ISP backbone, an enterprise gateway, or a hospital network — not only for a leaderboard position. The result is a practical selection framework: match the classifier to the constraints, and know what you are giving up when you do.