CommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition
summary
Automatically generates stealthy voice commands embedded into real songs so ASR decoders accept phrases while human listeners do not. Uses an open-source Kaldi/Howdy-style acoustic model plus a language-model/FST decoding attack pipeline to produce CommanderSong samples, models speaker and background noise, and validates over‑the‑air with JBL speakers at ~1.05 m and phones (iPhone, LG, Huawei). A 200‑person MTurk test finds essentially zero human recognition; attacks transfer to commercial ASR (iFlytek/phone apps) and via YouTube. Proposed mitigations include audio squeezing and added audio turbulence; authors report these reduce attack success.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:22Outline
- 2:01Motivation
- 3:24Commander Song Attack
- 3:59Challenges Of The Attack
- 7:59WTA Attack Approach
- 9:34WAA Attack Approach
- 10:22Evaluation
- 16:46Conclusion
- 22:59Defense
related talks
30mHidden Voice Commands
39m Put Words In My Mouth
33mYour voice confirms my identity
24mInaudible Voice Commands: The Long-Range Attack and Defense
48mSecurity Vulnerabilities in AI Assistant Based Applications
06mAnti-Spoofing Voice Commands: A Generic Wireless Assisted Design
14mPlease (Don't) Stop the Music: Adversarial Red-Teaming of AI...
06mCountering Acoustic Adversarial Attacks in Microphone-equipped Smart Home Devices