Frequency Domain Filtering for Robust Neural Networks

21 concepts16 papers1 proposal

Deep neural networks excel at finding patterns in spatial and temporal data, but they suffer from a fundamental vulnerability: spatial entanglement. In a standard image or time-series waveform, relevant structural features and irrelevant noise are tightly interwoven across the same pixels or time steps. If an attacker alters a few pixels, or if a sensor introduces environmental static, the distortion propagates through convolutional layers, corrupting the model's internal representations.

To bypass this limitation, a growing body of research shifts the analytical paradigm from the spatial/temporal domain to the frequency domain. By transforming data using tools like the Fast Fourier Transform (FFT), Discrete Cosine Transform (DCT), or Discrete Wavelet Transform (DWT), models can treat signals not as arrangements of coordinates, but as combinations of periodic rates of change.

This spectral perspective acts as a mathematical prism, separating overlapping phenomena into distinct, linearly separable bands. As a result, models can cleanly isolate, enhance, or discard specific components to achieve unprecedented robustness against noise, domain shifts, and adversarial attacks.


The Intuition: Why Noise Separates in the Frequency Domain

To understand why frequency-domain filtering works, consider how different types of information distribute across a spectrum. Natural data—whether it is a medical ultrasound, a satellite image, or a biological time-series—tends to be highly structured and mathematically sparse. Its core energy concentrates within specific, predictable frequency bands (often lower frequencies representing smooth shapes or distinct periodic rhythms).

Conversely, noise and corruptions typically exhibit entirely different spectral footprints:

  1. Adversarial Perturbations: An $L_0$ adversarial attack corrupts a small number of pixels with high-intensity alterations to fool a network. In the spatial domain, this looks like a localized spike. However, a fundamental property of the Fourier transform dictates that a localized spike in space spreads out thinly across the entire frequency spectrum. As demonstrated by Bafna et al., projecting a corrupted image onto its dominant frequency components (such as the top DCT coefficients) effectively strips away this distributed adversarial energy while leaving the core image structure intact.
  2. Stochastic and Sensor Noise: Environmental interference, such as speckle noise in Synthetic Aperture Radar (SAR) or sensor clutter in infrared tracking, typically manifests as high-frequency stochastic artifacts. In the spatial domain, this noise obscures target boundaries. In the frequency domain, it separates cleanly from the lower-frequency target signatures, allowing models to suppress the interference directly (Ji et al.; Xie et al.).

Mechanism 1: Spectral Masking for Adversarial Defense

Instead of relying on computationally expensive adversarial training, researchers are building robustness directly into network architectures by inserting frequency-based filtering layers before or within the feature extraction pipeline.

[Corrupted Input] ──> [Fourier/DCT Transform] ──> [Spectral Masking] ──> [Inverse Transform] ──> [Clean Features]
                                                         │
                                            (Drops high-frequency noise
                                             or distributed artifacts)

For instance, Xu et al. designed a defense for Spiking Neural Networks (SNNs) that applies a Discrete Fourier Transform to input images. By utilizing time-varying masks that prioritize low-frequency information, their framework selectively attends to core structural features while filtering out high-frequency adversarial perturbations.

Similarly, in the context of Image Quality Assessment (IQA), Chistyakova et al. introduced structural blocks that process signals in the frequency domain. This architectural choice leverages the unique spectral characteristics of adversarial perturbations to isolate and suppress them, allowing the model to withstand attacks naturally without data-driven defenses.


Mechanism 2: Decomposing Structural Signals from Environmental Clutter

Beyond defending against malicious attacks, frequency filtering is highly effective at isolating weak target signals from overwhelming background noise in real-world engineering and medical applications.

Medical Imaging and Generative Noise

In medical image segmentation, traditional spatial attention mechanisms often struggle with artifacts like blur or speckle noise. Lin et al. demonstrated that processing video frames in the frequency domain allows models to overcome these challenges because spectral changes capture global image structures more robustly than localized pixel operations.

This property is also vital in generative workflows. For example, Wu et al. integrated Fourier-domain filtering into a diffusion-based medical segmentation framework. By filtering features in the spectral domain, the authors successfully separated stable anatomical structural signals from the stochastic noise inherent to the iterative diffusion generation process.

Remote Sensing and Physical Diagnostics

For high-resolution satellite imagery, Moghalles et al. applied a 2D Discrete Cosine Transform to feature maps. This allowed the network to analyze multi-spectral frequency components, making it significantly easier to distinguish blurry building boundaries from complex background environmental noise.

The same principle applies to one-dimensional physical signals:

  • Acoustics: Sha et al. transformed industrial valve acoustic signals into the frequency domain, isolating the specific spectral bands of cavitation events from broadband background industrial noise.
  • Aerodynamics: Manohar et al. utilized spectral decomposition on wing strain time-series data, separating high-magnitude periodic inertial forces (the flapping of the wing) from the lower-magnitude external aerodynamic signals needed for flight environment identification.
  • Physiology: Spetlík et al. trained a spatial feature extractor by maximizing the signal-to-noise ratio of an output sequence relative to a ground-truth frequency, forcing the network to isolate subtle, periodic heart-rate signals directly from environmental and motion artifacts.

Mechanism 3: Inverting the Filter for Domain Generalization

While the most common use case involves discarding high frequencies to eliminate noise, the frequency domain can also be manipulated in reverse to achieve domain generalization.

When a model is trained on data from a single environment (e.g., clear day) and deployed in another (e.g., heavy rain or night), it often fails because it over-relies on low-frequency global features like color casts and lighting shifts.

Low Frequencies  (Color casts, lighting shifts) ──> DISCARD (Varies across domains)
High Frequencies (Edges, fine textures)         ──> KEEP    (Stable across domains)

To solve this, Li et al. inverted the traditional filtering logic for object detection. They utilized the frequency domain to extract high-frequency components—such as sharp edges and fine textures—while filtering out the highly variable low-frequency illumination components. Because structural edges remain stable across rain, fog, and darkness, this frequency-selective prompt engineering allows the model to generalize across drastically different environments.


Bypassing Spatial Reconstruction Entirely

The ultimate extension of this paradigm is to eliminate the spatial domain from the loop altogether. In traditional workflows, raw frequency data (like an MRI scan's raw $k$-space data) is first converted into a spatial image using an Inverse FFT, and then fed into a neural network. However, this reconstruction step can permanently bake in and propagate artifacts.

Huang et al. bypassed this entirely with SERANet, a framework that performs brain MRI segmentation directly on raw $k$-space frequency data. By operating entirely within the frequency domain, the model extracts features optimized purely for the downstream task rather than visual appeal, preventing the propagation of reconstruction errors and maintaining high performance even in severe noise conditions.

By treating the frequency domain as a primary computational space rather than a temporary diagnostic tool, modern architectures achieve a clean separation of signal and noise that spatial operations alone cannot replicate.

Key Foundational Work

Thwarting Adversarial Examples: An L_0-Robust Sparse Fourier Transform (2018) — contributed 2 concept elements to this direction.
Environment identification in flight using sparse approximation of wing strain (2017) — contributed 1 concept elements to this direction.
End-to-end Segmentation with Recurrent Attention Neural Network (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_frequency_domain_noise_filtering_7d47cd6b,
  title = {frequency-domain noise filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/frequency-domain-noise-filtering/rdr_7d47cd6b712a43c4832e9d8d968c34bf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_frequency_domain_noise_filtering_7d47cd6b_v1,
  title = {frequency-domain noise filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/frequency-domain-noise-filtering/rdr_7d47cd6b712a43c4832e9d8d968c34bf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake