Ghost in the Radio: An Audio Adversarial Attack Using Environmental Noise Through Radio

Abstract

Many automatic speech recognition (ASR) systems provide functions such as sending messages or opening garage doors. If such a system malfunctions, it could pose a significant threat. In this paper, we propose the attack scenario in which we transfer audio adversarial examples to radio system through the radio signal and the sound received through radio deceives the ASR system. In addition, we propose over-the-air technology using environmental noise so that we can attack easily and effectively. Using this method, we observed that the combined sound was much more robust to environmental noise than the non-combined sound. Finally, we experiment as our scenario on a real environment with adversarial examples created using environmental noise and show the results. Our attack succeeded against commercial ASR systems with a success rate of over 50%.

Keywords

• Adversarial attack
• artificial intelligence
• automatic speech recognition
• environmental noise
• over-the-air