Hamiltonian Transformer Boosts RF Fingerprinting Accuracy
The Hamiltonian Transformer, a novel deep learning framework informed by physics, enhances radio-frequency transmitter fingerprinting by implementing norm-preserving value dynamics through a learned skew-symmetric generator and Störmer-Verlet leapfrog integration. At the input layer, an extra phase-increment embedding reveals oscillator dynamics. When evaluated on non-equalized raw I/Q signals from the WiSig dataset, this model reaches an impressive accuracy of 99.12% for same-day classification and 61.64% when expanded to 150 transmitters, surpassing traditional transformers in handling distribution shifts.
Key facts
- Proposes Hamiltonian Transformer for RF fingerprinting
- Uses learned skew-symmetric generator and Störmer-Verlet integration
- Phase-increment embedding captures oscillator dynamics
- Tested on WiSig dataset with non-equalized raw I/Q signals
- Achieves 99.12% accuracy in same-day classification
- Achieves 61.64% accuracy at 150 transmitters
- Addresses receiver and channel distribution shifts
- Published on arXiv with ID 2605.30364
Entities
Institutions
- arXiv