Embedding Bits in the Over-the-Air Sounds: Threat and Countermeasure

聲音傳輸中的信息隱寫:威脅和對抗

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date7 Sep 2021

Abstract

More and more smart devices nowadays are equipped with microphones to provide a convenient device control and input interface through sound waves, e.g., via speech recognition, while the input sounds are believed to be trustworthy usually. In this dissertation, we observe the possibility of embedding some hidden bits into the sound waves. Such bits are inaudible to our ears, but the resulting sound waves can attack the speech recognition system on a smart device to recognize the received sound as any malicious transcript selected by the attacker. Motivated by this observation, we investigate the hidden bit embedding in acoustic signals systematically in this dissertation to unveil this real-world threat and also aim to develop an effective countermeasure. First, we observe that the added hidden bits can be distorted by the acoustic channel after a sound is transmitted over the air. We thus conduct extensive empirical measurement studies in various environments to understand the characteristics of the acoustic channel, where we obtain important insights to launch the consequent technical designs. Second, based on our measurement study, we develop a system to generate the imperceptible audio adversarial examples by adding carefully designed hidden bits that can survive over-the-air to attack the neural network of a speech recognition system. The key challenge stems from how to ensure the added hidden bits in advance at the sender side are immune to the unknown signal distortion during the transmission. We propose a set of techniques to demonstrate this possibility, which can alarm people to this potential real-world threat. Third, we further study the principle of the human auditory system and propose a new information embedding design by modulating hidden bits into the phase of acoustic signal. This design can enable a high-speed side-channel over the acoustic signals. Atop this side-channel, we propose an effective protocol to detect and avoid the over-the-air audio adversarial examples. With this side-channel, we are also able to enable novel room-scale connectivity-based services that are not viable by prior wireless protocols yet. In summary, this dissertation provides a comprehensive study for the hidden bit information added in the acoustic signals to unveil the potential threat, investigate the effective countermeasure and enable useful application designs.