Abstract
In this thesis we analyze the efficacy of Prototypical Neural Networks and Meta-learning in the problem space of radio frequency device fingerprint classification. Radio frequency
fingerprinting is used to differentiate radio emitters based on subtle defects in transmitter hardware. The device fingerprint is a subtle effect, which can easily be overwhelmed by environmental factors. Current state of the art efforts largely ignore these environment factors, and validate their approach only using single datasets. We apply Prototypical Neural Networks and Meta-Learning across multiple datasets (not just domains in a single dataset) in order to determine Prototypical Neural Networks’ ability to mitigate environmental effects during classification. We also analyze the efficacy of several pre-processing techniques. Data for this thesis was sourced from ORACLE, CORES, and WiSig, all of which are publicly available RF datasets. Our results definitively show that prototypical neural networks are effective for single dataset RF fingerprinting. We also show that prototypical neural networks may be effective for multi-dataset RF fingerprinting, though multi-dataset RF fingerprinting imposes significant challenges.