Voice AI Systems Are Vulnerable to Hidden Audio Attacks | Hacker News

Researchers say voice AI can be manipulated with hidden audio perturbations that change model outputs while staying hard for people to hear.

The attacks work by adding tiny input changes, creating adversarial examples that can fool speech and voice models.

Audio attacks pose unique challenges compared with images because they must balance imperceptibility and effectiveness in real-world sound.

The finding raises security concerns for voice-based AI systems such as speech recognition and other audio applications.