Low Frequency Adversarial Perturbations Bypass Common Defenses

44 concepts36 papers1 proposal

Adversarial attacks are often visualized as "pixel fuzz"—high-frequency jitter that is invisible to humans but catastrophic for Deep Neural Networks (DNNs). Early research, such as the work by Gu et al., suggested that these perturbations primarily reside in the high-frequency domain, leading to a wave of defenses designed to "purify" images by filtering out high-frequency noise.

However, a growing body of evidence suggests that this focus on high frequencies has created a false sense of security. By shifting adversarial signals into the low-frequency domain, attackers can bypass almost all standard defenses while creating perturbations that are more robust, more transferable, and harder to detect.

The High-Frequency Fallacy

The intuition behind many adversarial defenses is simple: if adversarial noise is high-frequency, we can remove it using low-pass filters like JPEG compression, Gaussian smoothing, or bit-depth reduction. Frank and Campbell demonstrate that this assumption is a critical vulnerability. They show that adversarial perturbations concentrated in low-frequency components are inherently robust against these transformations. Because these defenses are designed to preserve the "core" of the image (the low frequencies), they inadvertently protect the adversarial signal embedded within them.

Sharma et al. go further, characterizing many state-of-the-art defenses as "security theater." By constraining attacks to the low-frequency spectrum using the Discrete Cosine Transform (DCT), they found that perturbations could bypass defenses that were specifically optimized to stop high-frequency noise.

The Mechanism of Structural Exploitation

Why are low-frequency attacks so effective? The answer lies in how Convolutional Neural Networks (CNNs) process information. Anshumaan and Agarwal argue that CNNs rely heavily on global structural features—the "shapes" and "skeletons" of objects—which are encoded in low-frequency bands. By corrupting these structural approximations, an attacker can cause the model to lose its representation of the primary object entirely.

This sensitivity is exploited in several specialized domains:

  • 3D Vision: Stolik and Lang show that perturbing low-frequency spectral coefficients of 3D meshes tricks autoencoders into reconstructing entirely different shapes, while standard low-pass filters actually refine the adversarial signal by removing minor artifacts.
  • Time Series: Li et al. demonstrate that classification models often over-rely on local high-frequency features, but can be more fundamentally misled by perturbations that target the global signal.
  • Generative Models: Kassis et al. reveal that diffusion-based purification—a popular defense—is vulnerable to systemic low-frequency attacks because these perturbations do not resemble the high-frequency noise the diffusion models were trained to eliminate.

Transferability and Semantic Stealth

Low-frequency perturbations are not just harder to filter; they are also more "transferable." Guo et al. observe that because different neural network architectures share a fundamental reliance on large-scale structural features, a low-frequency attack designed for one model is highly likely to fool another.

Furthermore, these attacks can take the form of natural-looking semantic shifts. Shamsabadi et al. and Yuan et al. demonstrate "unrestricted" attacks that use smooth color shifts or realistic color distributions. These global changes bypass spectral filters because they mimic natural image properties, making them nearly imperceptible to humans while remaining highly effective against classifiers. Similarly, Guesmi et al. show that low-frequency perturbations like simulated translucent "raindrops" on a lens can bypass denoising filters that expect grainy pixel noise.

Evolving Robustness Strategies

To counter these threats, researchers are developing defenses that move beyond simple filtering. Pei and Ma propose FreqPure, which selectively preserves low-frequency amplitude and phase during diffusion to "wash away" adversarial noise without destroying the image's core energy. Li et al. introduce Dual Adversarial Training (DAT), which forces models to prioritize the "phase" of the frequency spectrum—where structural information lives—over stylistic amplitude patterns.

On the defensive side, Zhang et al. argue that models should be trained to focus on invariant structural components like overall shape. While this makes models more resilient to high-frequency noise, the ongoing challenge remains: as long as models rely on low-frequency features for recognition, those same channels remain a potent vector for adversarial manipulation. As Bibi and Hammoud note, when adversarial triggers overlap with primary recognition features, the statistical gap between clean and malicious samples narrows, making detection a fundamental hurdle for the field.

Key Foundational Work

Low Frequency Adversarial Perturbation (2018) — contributed 2 concept elements to this direction.
Towards Deep Neural Network Architectures Robust to Adversarial Examples (2015) — contributed 1 concept elements to this direction.
Transferable Adversarial Perturbations (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_low_frequency_adversarial_vulnerability_64d3383c,
  title = {low-frequency adversarial vulnerability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/low-frequency-adversarial-vulnerability/rdr_64d3383cab194c20b56f007c76307d8d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_low_frequency_adversarial_vulnerability_64d3383c_v1,
  title = {low-frequency adversarial vulnerability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/low-frequency-adversarial-vulnerability/rdr_64d3383cab194c20b56f007c76307d8d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake