Your voice confirms my identity
summary
Demonstrates that consumer-grade 11 Labs clones trained on 3–5 minutes or public videos can bypass commercial voice authentication, with a lab sample of 15 people showing a ~2% relaxation of acceptance threshold creates many false accepts and some generated clips reach ~80% success; attack requires only a laptop, speakers and a phone, and succeeded against a real customer account. Recommends layered defenses: MFA, user-chosen full-sentence passphrases, logging, and cryptographic provenance for generated audio.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:24Voice cloning attack demo
- 7:06Bypassing voice authentication
- 13:33Training a voice clone
- 20:11Cloning from digital presence
- 24:29Why voice isn't secure
- 27:19Mitigating voice authentication risks
- 31:12Recap and future outlook
- 32:12Closing remarks
related talks
39m Put Words In My Mouth
36mFabricated Realities: AI Personas and Simulacra and the Threats They Pose
39mSynthetic Trust: Exploiting Biases at Scale
06mVoice In Ear: Spoofing-Resistant and Passphrase-Independent Body Sound Authentication
06mVocalLock: Sensing Vocal Tract for Passphrase-Independent User Authentication Leveraging ...
35mThe UX Of Advanced Voice Biometrics For Authentication
45mVoice Security - Get Up to Speed on This Attack Vector
24mCommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition