Hidden Voice Commands
summary
Researchers generate audio that humans hear as noise but voice systems transcribe as commands, producing both a white-box inversion that backsolves target text through MFCC and gradient descent and a black-box audio obfuscator that fuzzes MFCC-space examples until APIs accept them. Attacks play over the air from speakers, control phones (examples: “OK Google take a picture,” “text 12345,” “browse to evil.com”), tolerate background noise in the black-box case, and defenses include a classifier achieving 99.8% detection accuracy.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:58The Problem
- 3:14Speech Recognition
- 5:31Whitebox Attack
- 9:46Threat Model
- 10:42Demo
- 12:55Limitations
- 13:23Solution
- 17:54Comparison
- 18:28Defenses
- 20:46Conclusion
- 21:36Questions
related talks
24mCommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition
39m Put Words In My Mouth
06mAnti-Spoofing Voice Commands: A Generic Wireless Assisted Design
24mInaudible Voice Commands: The Long-Range Attack and Defense
48mSecurity Vulnerabilities in AI Assistant Based Applications
45mAnalogy Cyber Security—From 0101 to Mixed Signals
06mCountering Acoustic Adversarial Attacks in Microphone-equipped Smart Home Devices
33mYour voice confirms my identity