IQ Data

IQ data consists of two orthogonal components: the In-Phase (I) component and the Quadrature (Q) component, separated by a 90-degree phase shift. This dual-channel representation preserves the instantaneous amplitude and phase of the signal, unlike simple power measurements.

• I component: Represents the real part, modulated by the cosine of the carrier wave
• Q component: Represents the imaginary part, modulated by the sine of the carrier wave
• Together they form a complex number (I + jQ) that fully captures the signal's state at each sample instant
• This structure enables direct extraction of hardware impairment fingerprints that manifest as subtle deviations from ideal constellation points

IQ data is generated by sampling the analog waveform at a rate that satisfies the Nyquist-Shannon sampling theorem, typically at least twice the signal bandwidth. This ensures lossless digital representation of all information contained within the band of interest.

• Sample rate determines the temporal resolution of captured hardware impairments
• Higher sample rates reveal transient phenomena during power amplifier ramp-up and ramp-down
• Oversampling beyond Nyquist provides additional data points for neural networks to learn subtle device-specific variations
• Modern SDRs capture IQ streams at rates from 20 MSPS to over 100 MSPS for wideband analysis

Raw IQ samples serve as the direct input tensor for deep learning models in RF fingerprinting applications. Unlike handcrafted features, neural networks learn to extract discriminative patterns directly from the complex-valued time series.

• 1D CNNs process IQ sequences as multi-channel time series, learning temporal filters that detect hardware-specific distortions
• Complex-valued neural networks operate directly on I and Q without separating them, preserving phase relationships
• Input tensors typically shaped as [batch_size, 2, num_samples] where the 2 channels represent I and Q
• Raw IQ eliminates information loss from intermediate transformations like spectrogram generation

The complex nature of IQ data explicitly encodes both magnitude (signal envelope) and phase (angular position) at every sample point. Hardware impairments manifest as distinct, measurable deviations in both dimensions.

• I/Q imbalance appears as asymmetry between the I and Q channel gains and phase orthogonality
• Phase noise from local oscillator imperfections creates characteristic jitter patterns visible in the Q component
• AM-AM and AM-PM distortion from power amplifier non-linearity alters both magnitude and phase relationships
• These impairments form a unique, unclonable signature that persists across different transmitted data payloads

IQ data generates substantial data volumes, with a single second of capture at 50 MSPS producing 100 million complex samples. This high dimensionality provides rich information for deep learning but demands efficient processing pipelines.

• A single IQ capture for fingerprinting typically spans 1,024 to 65,536 samples per burst
• Storage requirements: each complex sample occupies 4-8 bytes (16-bit or 32-bit precision per component)
• Dimensionality reduction techniques like PCA or autoencoders compress IQ data while preserving fingerprint features
• Edge deployment requires model compression and optimized inference to process IQ streams in real-time

Raw IQ data inherently contains both the transmitter-intrinsic impairments used for fingerprinting and the channel-induced distortions that must be overcome. Robust deep learning models learn to disentangle these superimposed effects.

• Multipath fading introduces frequency-selective distortion that varies with device location
• Doppler shift from relative motion alters the apparent carrier frequency
• Additive white Gaussian noise (AWGN) obscures subtle hardware signatures at low SNR
• Domain adaptation and contrastive learning techniques train models to isolate device-specific features from channel effects
• Data augmentation with synthetic channel models improves generalization to unseen environments