Wavelet Scattering Transform

The WST achieves translation invariance through a cascade of wavelet convolutions, pointwise complex modulus operations, and local averaging with a low-pass filter. This structure ensures that small time shifts in the input signal do not alter the final representation, a critical property for RF fingerprinting where signal alignment is never perfect. Unlike the Fourier transform, which loses all temporal information, the scattering transform preserves high-frequency details while building invariance layer by layer.

A defining property of the WST is Lipschitz stability to small diffeomorphisms. This means that minor, non-linear distortions of the signal—such as those caused by multipath fading or Doppler shifts—produce proportionally small changes in the scattering coefficients. This stability is mathematically guaranteed by the wavelet filter design and is essential for ensuring that a device fingerprint remains consistent across varying channel conditions.

The transform computes a hierarchical tree of coefficients by iterating wavelet convolutions and modulus operations. Each layer captures interactions across increasing time scales and frequency bands:

• Layer 1 (S1): Captures local spectral energy and transient events.
• Layer 2 (S2): Captures modulation spectra and interactions between frequency components, such as those generated by amplifier non-linearity. This multi-scale analysis naturally separates transient and steady-state features.

Unlike learned deep neural networks, the WST uses fixed wavelet filters that are not trained via backpropagation. The filters are designed based on physical principles, such as constant-Q frequency spacing, which mirrors the logarithmic frequency resolution of the human auditory system and RF signal structures. This eliminates the need for large labeled datasets for feature extraction, making it highly data-efficient and interpretable.

The scattering transform is designed to be a contractive operator that conserves signal energy as it propagates through the network. The energy of the input signal is decomposed into the scattering coefficients, and the energy at each layer decays exponentially. This property ensures that the representation is not only stable but also preserves the discriminative information necessary to distinguish between transmitters with subtle hardware variations.

The modulus operator and subsequent averaging provide inherent robustness to additive Gaussian noise. High-frequency noise components are isolated in the first layer of wavelet coefficients and are significantly attenuated by the low-pass averaging filter. This makes the WST particularly effective for extracting clean device fingerprints from low-SNR signals, a common challenge in wideband spectrum monitoring and long-range IoT authentication.

The foundational domain for wavelet scattering, where signals are analyzed jointly in time and frequency. Unlike the Short-Time Fourier Transform, which uses a fixed resolution, wavelet scattering provides a multi-resolution decomposition that captures both transient events and steady-state oscillations. This is critical for isolating non-stationary hardware impairments like amplifier memory effects and phase noise that manifest differently across time scales.

A core property of the scattering transform achieved through wavelet modulus operators and low-pass averaging. For RF fingerprinting, this means the representation remains stable even when a signal burst is shifted in time due to variable propagation delays or packet detection jitter. This eliminates the need for precise temporal alignment before feature extraction, making the system robust to real-world capture imperfections.

Wavelet scattering provides Lipschitz continuity to small diffeomorphisms—meaning minor warping or stretching of the signal does not catastrophically change the representation. In RF contexts, this translates to resilience against Doppler shifts, clock drift, and subtle channel distortion. The scattering coefficients encode the signal's geometric structure in a way that is locally stable but globally discriminative.

Beyond first-order wavelet modulus coefficients, the scattering transform cascades operations to capture interference patterns and non-linear interactions between frequency components. Second-order coefficients encode the amplitude modulation of one frequency band by another, revealing subtle cross-modulation products and intermodulation distortion that are highly specific to individual transmitter hardware chains.

A classic feature extraction method originally for speech, MFCCs apply a mel-scale filterbank and discrete cosine transform to capture spectral envelope shape. While computationally efficient, MFCCs lack the deformation stability and multi-scale structure of wavelet scattering. In RF fingerprinting, scattering transforms often outperform MFCCs when signals exhibit complex, non-stationary hardware impairments across multiple time scales.

The wavelet scattering transform can be viewed as a pre-defined, non-learned convolutional network with fixed wavelet filters, modulus non-linearities, and pooling operations. This architecture mirrors the hierarchical feature extraction of CNNs but requires no training data. For RF fingerprinting with limited labeled samples, this provides a powerful, mathematically grounded feature extractor that avoids overfitting while capturing the hierarchical structure of hardware impairments.