eess.ASMar 17, 2026
SaveOver-the-air White-box Attack on the Wav2Vec Speech Recognition Neural Network
Organizations: Joint Stock Research and Production Company Kryptonite
Abstract
Automatic speech recognition systems based on neural networks are vulnerable to adversarial attacks that alter transcriptions in a malicious way. Recent works in this field have focused on making attacks work in over-the-air scenarios, however such attacks are typically detectable by human hearing, limiting their potential applications. In the present work we explore different approaches of making over-the-air attacks less detectable, as well as the impact these approaches have on the attacks' effectiveness.
Figures & tables
Figure 1: Data augmentation.
Figure 2: Example room layout.
Figure 3: Attack generation iteration. Gradient backpropagation is indicated by red arrows.
Figure 4: Audio fragments padded with zeros.
| Experiment | FR | RS | TS | Success rate | PER | WER | Generation time, hrs | |
|---|---|---|---|---|---|---|---|---|
| 1 | + | + | + | 0.00 | 0% | 0.13 | 0.44 | 1.0 |
| 2 | + | + | + | 0.15 | 40% | 0.03 | 0.09 | 11.5 |
| 3 | — | + | + | 0.00 | 0% | 0.20 | 0.45 | 1.4 |
| 4 | — | + | + | 0.15 | 20% | 0.06 | 0.20 | 8.2 |
| 5 | + | — | + | 0.00 | 0% | 0.94 | 0.86 | 0.3 |
| 6 | + | — | + | 0.15 | 0% | 0.91 | 0.87 | 1.1 |
Table 1: Experimental results. FR — frequency response, RS — room simulation, TS — time shift.
Figure 5: Numbers of iterations required to reach various values of .
More in this topic
Adversarial Examples
Adversarial examples are inputs crafted to cause machine-learning models to misclassify, misperceive, or generate unsafe outputs.
See all 76 papers