Overview
Explore the fascinating world of audio adversarial examples in this 25-minute IEEE conference talk presented at the 1st Deep Learning and Security Workshop. Delve into the construction of targeted attacks on automatic speech recognition systems, specifically focusing on Mozilla's DeepSpeech implementation. Discover how audio waveforms can be manipulated to be 99.9% similar to the original while transcribing as any chosen phrase, with the ability to recognize up to 50 characters per second. Gain insights into the white-box iterative optimization-based attack methodology and its 100% success rate. Understand the implications of this new domain in the study of adversarial examples and its potential impact on speech-to-text technology security.
Syllabus
Audio Adversarial Examples: Targeted Attacks on Speech-to-Text
Taught by
IEEE Symposium on Security and Privacy