Psychoacoustic Masking Thresholds for Adversarial Audio Perturbations

This direction is merged into another direction: over-the-air robustness. Its last version stays readable and citable here.

Version 1 of 113 papers16 conceptsfrozen 2026-10-04

In the world of computer vision, adversarial attacks are often measured by $L_p$ norms—mathematical constraints that ensure the "distance" between an original image and a perturbed one is small enough to be invisible. However, as researchers like Sajjad Abdoli and Patrick Cardinal have noted in Universal Adversarial Audio Perturbations, these metrics fail spectacularly when applied to sound. A mathematically "small" perturbation in a waveform can manifest as loud, jarring static to a human listener, even if it successfully tricks a machine learning model.

The emerging solution to this problem is to move away from pure mathematics and toward psychoacoustics: the study of how humans actually perceive sound. By exploiting the biological limits of the human ear, researchers are creating adversarial audio that is effectively silent to people but remains "loud" and clear to Automatic Speech Recognition (ASR) systems.

The Mechanism of Auditory Masking

The core intuition behind these attacks is the masking threshold. Human hearing is not a linear recorder; it is a selective filter. When a loud sound (the "masker") occurs at a specific frequency, the human ear becomes temporarily unable to hear quieter sounds (the "maskee") at nearby frequencies.

In Imperceptible, Robust, and Targeted Adversarial Examples for Automatic Speech Recognition, Yao Qin and Nicholas Carlini demonstrated that you can hide a malicious signal "underneath" the frequency of the original audio. Instead of trying to keep the total energy of the noise low (as an $L_p$ norm would), they optimize the noise so that its power spectrum always stays below the masking threshold of the source audio.

This allows for perturbations with significantly higher energy than traditional methods, which makes the attacks more robust. Because the noise is "hidden" by the existing audio content, it remains imperceptible to humans even if the volume is turned up.

From Stealth to Robustness

A major challenge in audio attacks is the "over-the-air" problem. Most adversarial examples are designed for digital files; once played through a speaker and recorded by a microphone in a real room, the echoes and distortions (reverberation) usually destroy the delicate adversarial signal.

However, by leveraging psychoacoustic principles, researchers have found ways to make attacks survive the physical world. Yao Qin and colleagues achieved this by incorporating a room impulse response (RIR) simulator into their optimization loop. By ensuring the adversarial signal remains below the masking threshold even after accounting for room reflections, they created the first targeted attacks that remain effective and imperceptible in physical environments.

Similarly, Qing Wang 0039 and Lei Xie 0001 applied these concepts to speaker recognition systems in Inaudible Adversarial Perturbations for Targeted Attack in Speaker Recognition. They achieved a 98.5% success rate by hiding malicious signals beneath the threshold of human hearing, proving that the "faults" in our ASR systems are not just digital artifacts but fundamental gaps between machine processing and human biology.

Why Psychoacoustics Matters for Security

The shift toward psychoacoustic modeling represents a maturation of the field. Early attacks were easy to detect because they sounded like "noise." As Hadi Abdullah and Patrick Traynor argue in their systematization of knowledge, traditional distance metrics are insufficient for evaluating the stealth of an attack.

The research also draws on older psychoacoustic findings, such as the work by Bert De Coensel and Dick B.M. Botteldooren on noise annoyance. Their study of high-speed trains showed that "annoyance" is driven by spectral distribution and "rise time" (how fast a sound reaches its peak) rather than just raw decibel levels. This reinforces the idea that for an adversarial attack to be truly "stealthy," it must respect the complex, non-linear ways the human brain processes temporal and spectral information.

By integrating these biological constraints, researchers are moving toward a more realistic understanding of the vulnerabilities in voice-controlled systems—vulnerabilities that exist precisely because machines "hear" the parts of the signal that humans have evolved to ignore.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_psychoacoustic_masking_thresholds_for_ad_7d8d6df6,
  title = {Psychoacoustic Masking Thresholds for Adversarial Audio Perturbations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/psychoacoustic-masking-thresholds-for-adversarial-audio-perturbations/rdr_7d8d6df69dc944e1914f1a5379b43b98},
  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_psychoacoustic_masking_thresholds_for_ad_7d8d6df6_v1,
  title = {Psychoacoustic Masking Thresholds for Adversarial Audio Perturbations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/psychoacoustic-masking-thresholds-for-adversarial-audio-perturbations/rdr_7d8d6df69dc944e1914f1a5379b43b98/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}