AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

Are Deep Speech Denoising Models Robust to Adversarial Noise?

Will Schwarzer, Andrea Fanelli, Philip S. Thomas, Xiaoyu Liu

ICLR 2026 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use full psychoacoustic masking (offset -12 dB, temporal masking) vs frequency-only masking (ablation)masked PGD attack on Demucs · 30 dB SNR + reverballowed masking offset at equal attack success: -8.4 dB -> -12 dBImproved
remove the -12 dB threshold shift (original masking thresholds) (ablation)masked PGD attack · 30 dB SNR + reverbattack strength (Delta STOI): -1.1 -> -1.4Improved
apply imperceptible adversarial noise to DNS model inputDemucs, FRCRN, FSN+, MP-SENet · DNS Challenge speechDelta STOI: initial +0.031..+0.057 -> -0.338..-1.258Hurt
transfer attack crafted on one DNS architecture to anothermasked PGD attack · cross-model transferDelta STOI: off-diagonal 0.03 to 0.08 vs on-diagonal -0.49 to -1.25No clear effect
train attack on multiple Demucs checkpoints and test on held-out onemasked PGD attack · Demucs leave-one-outspeech qualityMixed
use universal perturbation instead of per-utterancemasked PGD attack · DNS modelsoutput degradationNo clear effect
targeted attack toward same-speaker utterancemasked PGD attack · DNS modelsSTOI to target vs intelligibility: STOI to target above 0.5Mixed
add Gaussian noise defense to attacked audioDNS under attack · DNS modelsDelta STOIMixed
simulate over-the-air attack with RIR on perturbationmasked PGD attack · DNS models except FSN+Delta STOINo clear effect
tighten masking offset from -8 dB to -16 dB (ablation)masked PGD attack · averaged over four modelsattack successHurt
FSN+ exploding gradients (gradient norm 1e30)FSN+ · white-box attackattack success: Delta STOI -0.338 vs -1.034..-1.258 for othersImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy