Multi-Class Device Identification on Large-Scale WiFi Network Traffic Data

We present a large-scale device identification study for six classes using over 110 million anonymized time series WiFi traffic data points from 11,513 unique devices. Through an ablation consisting of 44 ensemble machine learning models with differing modeling and featurization approaches, we demonstrate that Random Forest classifiers leveraging rich temporal features achieve strong multi-class performance (F1-macro: 0.77 validation, 0.68 test), from which the device’s transmission throughput and received signal strength features emerge as most important.