SurfingAttack: Interactive Hidden Attack on Voice Assistants Using Ultrasonic Guided Waves

Abstract

With recent advances in artificial intelligence and natural language processing, voice has become a primary method for human-computer interaction. It has enabled game-changing new technologies in both commercial sectors and military sectors, such as Siri, Alexa, Google Assistant, and voice-controlled naval warships. Recently, researchers have demonstrated that these voice assistant systems are susceptible to signal injection at the inaudible frequencies. To date, most of the existing works focus primarily on delivering a single command via line-of-sight ultrasound speaker or extending the range of this attack via speaker array. However, besides air, sound waves also propagate through other materials where vibration is possible. In this work, we aim to understand the characteristics of this new genre of attack in the context of different transmission media. Furthermore, by leveraging the unique properties of acoustic transmission in solid materials, we design a new attack called SurfingAttack that would enable multiple rounds of interactions between the voice-controlled device and the attacker over a longer distance and without the need to be in line-of-sight. By completing the interaction loop of inaudible sound attack, SurfingAttack enables new attack scenarios, such as hijacking a mobile Short Message Service (SMS) passcode, making ghost fraud calls without owners' knowledge, etc. To accomplish SurfingAttack, we have solved several major challenges. First, the signal has been specially designed to allow omni-directional transmission for performing effective attacks over a solid medium. Second, the new attack enables multi-round interaction without alerting the legitimate user at the scene, which is challenging since the device is designed to interact with users in physical proximity rather than sensors. To mitigate this newly discovered threat, we also provide discussions and experimental results on potential countermeasures to defend against this new threat.

Measuring the Effectiveness of Privacy Policies for Voice Assistant Applications

“…One line of research is to exploit interpretation errors of user commands by speech recognition, such as voice squatting attack [30,47], and generate hidden/inaudible voice commands [21,35,36,40,42,44,46]. Another line of research focuses on defense mechanisms, including, continuous authentication [25], canceling unwanted baseband signals [46], correlating magnetic changes with voice commands [22], and user presence-based access control [32].…”

Smart Home Voice Assistants: A Literature Survey of User Privacy and Security Vulnerabilities

“…For instance, lasers can be used to activate the devices and allow injecting simulated voice commands into the device's internal microphone [18]. Similarly, researchers have found ways to bypass device passcode restrictions on smart phones and inject voice commands into the voice assistant software using "guided ultrasonic waves," which allowed them to place unauthorized phone calls on the target device [19]. Other attacks may target the cloud ecosystem of the IVA [3].…”

“…We filtered the results following the procedure in [19] by applying the inclusion and exclusion criteria from Section 3.3. Search engine result pages were manually parsed and studies were selected based on their titles.…”

“…In [3], Dolphin attacks are highlighted as a variation of a replay attack. Dolphin attacks could occur when commands are delivered through ultrasonic frequencies, much like the SurfingAttack presented in [19]. Dolphin attacks exploit the vulnerability of the off-the-shelf microphone components used in some IVA devices, which are able to capture frequencies outside of the audible range, allowing an attacker to gain control of IVA devices.…”

“…However, since these attacks make use of synthesized high-frequency sounds, the required close proximity of the attacker to the target device and the required special hardware make it difficult to perform in a real-world setting, according to [3]. Yet, as [19] shows, these attacks can be used to take control over the user's device and perform unauthorized actions, such as placing phone calls, and thereby possibly submit personal data to a hacker-controlled medium.…”