Hamiltonian-Inspired Attention Mechanism for Scalable RF Transmitter Fingerprinting

Researchers propose the Hamiltonian Transformer, a physics-informed deep learning architecture for identifying wireless transmitters via RF fingerprinting that achieves 99.12% accuracy in controlled settings but maintains 61.64% accuracy when scaling to 150 devices. The model uses norm-preserving attention mechanisms inspired by Hamiltonian mechanics to improve generalization across receiver types, channels, and time periods compared to standard CNN and Transformer baselines.

This research addresses a critical challenge in wireless security and device authentication: the degradation of deep learning models when deployed at scale or under distribution shifts. RF fingerprinting—identifying transmitters by their hardware-induced signal imperfections—has become increasingly important for network security, but existing approaches falter as transmitter populations grow and operating conditions change. The Hamiltonian Transformer introduces physics-informed structural priors into attention mechanisms, leveraging conservation laws from Hamiltonian mechanics to enforce norm preservation in attention value dynamics.

The work builds on emerging trends in physics-informed machine learning, where domain knowledge constraints guide neural network architectures toward more robust and generalizable solutions. By embedding oscillator dynamics at the input layer and using skew-symmetric generators with Störmer-Verlet integration, the model captures underlying signal characteristics more effectively than purely data-driven approaches. The ablation study reveals that norm-preservation in value updates drives the scaling advantage, while phase-increment embeddings provide substantial per-component improvements.

For wireless security infrastructure and network operators, this approach offers practical implications: improved device authentication at scale without relying on channel equalization or controlled experimental conditions. The scalability advantage—maintaining reasonable accuracy at 150 transmitters—matters for deployed systems managing hundreds or thousands of devices. Telecommunications standards bodies and security researchers monitoring wireless authentication methods should track adoption of physics-informed architectures, as they represent a methodological shift toward more interpretable and robust machine learning models in signal processing applications.

• →Hamiltonian Transformer achieves 99.12% same-day accuracy but degrades to 61.64% at 150 transmitters, showing scalability challenges persist despite improvements.
• →Physics-informed attention mechanisms with norm-preserving dynamics outperform standard transformers and CNNs across all tested distribution shifts.
• →The model operates on raw, non-equalized I/Q signals, eliminating preprocessing steps and improving practical deployment viability.
• →Ablation studies identify norm-preservation as the primary scaling advantage, validating the use of physics constraints in neural architecture design.
• →Cross-receiver and cross-day generalization results demonstrate robustness to environmental variations critical for real-world wireless authentication systems.