Hear "No Evil", See "Kenansville" - Efficient and Transferable Black-Box Attacks on Automatic Speech Recognition Systems

Explore efficient and transferable black-box attacks on automatic speech recognition systems in this 14-minute IEEE conference presentation. Delve into various attack types against machine learning models, including untargeted and targeted attacks, as well as generic attacks. Examine the voice system pipeline and understand how these attacks function, observing their impact on transcription and tag capabilities. Watch a demonstration and learn about user study findings. Gain valuable insights into the vulnerabilities of speech recognition technology and potential security implications in this comprehensive overview.

Introduction
Attacks against ML Models
Untargeted Attacks
Target Attacks
Generic Attack
Voice System Pipeline
How does the attack work
Observation
Impact on Transcription
Tag Capabilities
Demo
User Study
Summary