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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use full psychoacoustic masking (offset -12 dB, temporal masking) vs frequency-only masking (ablation) | masked PGD attack on Demucs · 30 dB SNR + reverb | allowed masking offset at equal attack success: -8.4 dB -> -12 dB | Improved |
| remove the -12 dB threshold shift (original masking thresholds) (ablation) | masked PGD attack · 30 dB SNR + reverb | attack strength (Delta STOI): -1.1 -> -1.4 | Improved |
| apply imperceptible adversarial noise to DNS model input | Demucs, FRCRN, FSN+, MP-SENet · DNS Challenge speech | Delta STOI: initial +0.031..+0.057 -> -0.338..-1.258 | Hurt |
| transfer attack crafted on one DNS architecture to another | masked PGD attack · cross-model transfer | Delta STOI: off-diagonal 0.03 to 0.08 vs on-diagonal -0.49 to -1.25 | No clear effect |
| train attack on multiple Demucs checkpoints and test on held-out one | masked PGD attack · Demucs leave-one-out | speech quality | Mixed |
| use universal perturbation instead of per-utterance | masked PGD attack · DNS models | output degradation | No clear effect |
| targeted attack toward same-speaker utterance | masked PGD attack · DNS models | STOI to target vs intelligibility: STOI to target above 0.5 | Mixed |
| add Gaussian noise defense to attacked audio | DNS under attack · DNS models | Delta STOI | Mixed |
| simulate over-the-air attack with RIR on perturbation | masked PGD attack · DNS models except FSN+ | Delta STOI | No clear effect |
| tighten masking offset from -8 dB to -16 dB (ablation) | masked PGD attack · averaged over four models | attack success | Hurt |
| FSN+ exploding gradients (gradient norm 1e30) | FSN+ · white-box attack | attack success: Delta STOI -0.338 vs -1.034..-1.258 for others | Improved |
Closest papers by meaning
- Defending against Adversarial Audio via Diffusion Model (ICLR 2023 · similarity 0.87)
- An Introduction to Adversarially Robust Deep Learning (TPAMI 2023 · similarity 0.87)
- Removing Adversarial Noise in Class Activation Feature Space (ICCV 2021 · similarity 0.86)
- AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language Models (ICLR 2025 · similarity 0.86)
- First Line of Defense: A Robust First Layer Mitigates Adversarial Attacks (AAAI 2025 · similarity 0.86)
- Breaking Adversarial Transferability in Fine-Tuned Speech Recognition (NeurIPS 2026 · similarity 0.85)
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.
