Research map / Adversarial robustness and security / Attacks on text and graphs
Imperceptible, Robust, and Targeted Adversarial Examples for Automatic Speech Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Constrain the perturbation using the audio masking threshold | Targeted Lingvo attack on LibriSpeech test-clean · LibriSpeech test-clean, no room simulation | sentence-level accuracy: 58.60 -> 100.00 | Helped |
| Optimize the perturbation against sampled room reverberations | Targeted Lingvo attack · LibriSpeech test-clean, simulated test rooms | sentence-level accuracy: 0% -> 62.96 | Helped |
| Increase the perturbation bound from Δ = 300 to Δ = 400 (ablation) | Robust Lingvo attack · LibriSpeech test-clean, simulated test rooms | sentence-level accuracy: 62.96 -> 64.64 | Helped |
| Increase the perturbation bound from Δ = 300 to Δ = 400 (ablation) | Robust Lingvo attack · LibriSpeech test-clean, simulated test rooms | WER: 14.45 -> 13.83 | Helped |
| Add the masking-threshold loss to the robust attack (ablation) | Robust Lingvo attack · LibriSpeech test-clean, simulated test rooms | sentence-level accuracy: 62.96 -> 49.65 | Hurt |
| Add the masking-threshold loss to the robust attack (ablation) | Robust Lingvo attack · LibriSpeech test-clean, simulated test rooms | WER: 14.45 -> 22.98 | Hurt |
| Use masking-threshold perturbations instead of the prior attack's perturbations | Targeted ASR adversarial examples · Human study: clean/noisy judgment | participants judging audio noisy: 66% -> 23% | Helped |
| Use masking-threshold perturbations instead of the prior attack's perturbations | Targeted ASR adversarial examples · Human study: identical-or-not judgment | judged identical to clean audio: 24% -> 76% | Helped |
| Fine-tune robust examples with the masking-threshold loss (ablation) | Robust Lingvo attack · Human study: clean versus adversarial audio, original identification | correctly identified original audio: 97% -> 89% | Helped |
Closest papers by meaning
- Characterizing Audio Adversarial Examples Using Temporal Dependency (ICLR 2019 · similarity 0.91)
- Weighted-Sampling Audio Adversarial Example Attack (AAAI 2020 · similarity 0.91)
- Rethinking Audio-Visual Adversarial Vulnerability from Temporal and Modality Perspectives (ICLR 2025 · similarity 0.87)
- Breaking Adversarial Transferability in Fine-Tuned Speech Recognition (NeurIPS 2026 · similarity 0.86)
- AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language Models (ICLR 2025 · similarity 0.85)
- Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise (AAAI 2024 · similarity 0.84)
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.