Inaudible Manipulation of Voice-Enabled Devices Through BackDoor Using Robust Adversarial Audio Attacks: Invited Paper

Abstract

The BackDoor system provides a method for inaudibly transmitting messages that are recorded by unmodified receiver microphones as if they were transmitted audibly. Adversarial Audio attacks allow for an audio sample to sound like one message but be transcribed by a speech processing neural network as a different message. This study investigates the potential applications of Adversarial Audio through the BackDoor system to manipulate voice-enabled devices, or VEDs, without detection by humans or other nearby microphones. We discreetly transmit voice commands by applying robust, noise-resistant adversarial audio perturbations through BackDoor on top of a predetermined speech or music base sample to achieve a desired target transcription. Our analysis compares differing base carriers, target phrases, and perturbation strengths for maximal effectiveness through BackDoor. We determined that such an attack is feasible and that the desired adversarial properties of the audio sample are maintained even when transmitted through BackDoor.