SpeechGuard: Recoverable and Customizable Speech Privacy Protection...
summary
System applies a reversible multi-parameter warping function (invertible voice conversion with randomized per-segment alpha/beta) plus automatic or manual sensitive-text detection and per-frame encryption, enabling frame-level reversible protection where warping parameters and encryption keys are assigned to permission groups (L1/L2/L3) and delivered via attribute-based encryption. Distortion strength is tuned (recommended ~0.5–0.6) to raise ASV error while preserving ASR accuracy. Evaluation on three datasets outperforms three baselines in anonymity, confidentiality, and attack resistance.
AI-written summary · about
chapters (13)
- 0:00SpeechGuard project intro
- 2:21Privacy risks in speech data
- 4:29Current methods' limits
- 5:11SpeechGuard goals
- 6:38Listener permissions
- 8:25System overview
- 10:30Voice conversion for privacy
- 13:05Privacy versus speech quality
- 14:31Sensitive text detection
- 15:58Frame-level access control
- 17:06Evaluation and strengths
- 18:40Limitations and future work
- 19:24Conclusion
related talks
06mVoice In Ear: Spoofing-Resistant and Passphrase-Independent Body Sound Authentication
30mHidden Voice Commands
06mVocalLock: Sensing Vocal Tract for Passphrase-Independent User Authentication Leveraging ...
24mCommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition
06mAnti-Spoofing Voice Commands: A Generic Wireless Assisted Design
33mYour voice confirms my identity
06mCountering Acoustic Adversarial Attacks in Microphone-equipped Smart Home Devices
45mVoice Security - Get Up to Speed on This Attack Vector