Domain Randomization

The systematic variation of simulator inputs across predefined uniform or log-uniform distributions to cover the full operational envelope. Key parameters include:

• SNR range: -10 dB to 30 dB
• Carrier Frequency Offset: ±10 ppm of center frequency
• Phase noise masks: -80 to -120 dBc/Hz at 10 kHz offset
• Multipath tap counts: 1 to 12 resolvable paths

The goal is to expose the model to combinations of impairments that may never occur together in real captures, forcing it to disentangle correlated features.

By randomizing non-identity-bearing parameters like channel impulse response and additive noise, the neural network is forced to ignore these nuisance variables. The model learns to latch onto the only consistent signal across all randomized environments: the hardware-specific impairment fingerprint.

This process implicitly performs a form of contrastive learning, where different randomized views of the same transmitter are pulled together in embedding space while views of different transmitters are pushed apart.

A core challenge in sim-to-real transfer is the domain gap between synthetic and real RF data. Domain randomization bridges this gap by making the simulation so broad that real-world conditions appear as just another sample from the training distribution.

• Extreme randomization: Apply impairment levels beyond physically plausible ranges
• Random latency: Jitter the timing of synthetic impairments
• Background interference: Inject random modulated interferers

This 'brute force' approach trades simulation fidelity for distributional coverage.

Rather than applying uniform randomization throughout training, a curriculum strategy progressively increases the difficulty and breadth of randomization:

1. Phase 1: Train on clean signals with only AWGN variation
2. Phase 2: Introduce mild multipath and CFO randomization
3. Phase 3: Add severe non-linear PA distortion and phase noise
4. Phase 4: Full randomization across all parameters simultaneously

This staged approach prevents the model from collapsing to a degenerate solution early in training when the task is too difficult.

An advanced variant where a second neural network learns to generate the most challenging randomization parameters for the primary fingerprinting model. This creates a min-max game:

• Generator: Searches for simulation parameters that maximize the fingerprinting model's loss
• Fingerprinter: Learns features robust to these worst-case perturbations

This approach automatically discovers failure modes in the randomization space that hand-tuned distributions might miss, such as specific combinations of AM-PM distortion and Doppler spread that are particularly deceptive.

Not all randomization is beneficial. Excessive randomization can destroy the very hardware signatures the model needs to learn. A randomization budget defines the maximum allowable perturbation before the device identity is obscured:

• I/Q imbalance: Randomize within ±2 dB gain, ±5° phase
• PA non-linearity: Vary AM-AM knee point by ±3 dB
• Carrier leakage: Randomize DC offset within -30 to -15 dBc

Parameters must respect the physical constraints of real hardware—randomizing beyond these bounds creates unrealistic signals that harm rather than help generalization.