In plain words: By nudging an audio waveform step by step while checking what the speech-to-text system would transcribe, they turn any recording into one that reads out a chosen phrase. The tweaked audio stayed over 99.9% similar to the original yet fooled the system every time.
Abstract · Audio Adversarial Examples: Targeted Attacks on Speech-to-Text
We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduce a new domain to study adversarial examples.
Nicholas Carlini, David Wagner
arXiv:1801.01944 · cs.LG, cs.AI, cs.CR · submitted Jan 5, 2018 · updated Mar 30, 2018
abstract · pdf · html
