Voice AI Systems Are Vulnerable to Hidden Audio Attacks
A study published in IEEE Spectrum reveals that voice AI systems are susceptible to adversarial audio attacks. Researchers from the University of California, Berkeley, and other institutions developed a technique to embed hidden commands into audio files that are inaudible to humans but recognized by voice assistants like Amazon Alexa, Google Assistant, and Apple Siri. The attacks exploit the way these systems process audio, using psychoacoustic masking to hide commands within benign sounds. In tests, the hidden commands successfully triggered actions such as dialing phone numbers, opening websites, and even unlocking smart locks. The researchers note that current defenses, like voice biometrics, are insufficient against these attacks. They recommend that developers implement more robust audio processing techniques and user verification methods to mitigate the risk. The findings highlight a growing security concern as voice AI becomes more integrated into daily life.
Developers must harden voice AI systems against imperceptible audio attacks to prevent unauthorized actions.