SurfingAttack: Interactive Hidden Attack on Voice
Assistants Using Ultrasonic Guided Waves
Qiben Yan
∗†
, Kehai Liu
∗‡
, Qin Zhou
§
, Hanqing Guo
†
, Ning Zhang
¶
†
Computer Science & Engineering, Michigan State University, {qyan, guohanqi}@msu.edu
‡
Songshan Lake Materials Laboratory, Institute of Physics, Chinese Academy of Sciences, liukehai@sslab.org.cn
§
¶
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.
I. INTRODUCTION
Recent advances in artificial intelligence (AI) and machine
learning have enabled new game-changing technologies for
humans to interact with machines. Conversation with AI is
no longer a scene in science-fiction movies, but day-to-day
routines. It is now possible for everyday users to converse
∗
The first two authors contributed equally to this work. Dr. Liu conducted
this work when he was a postdoctoral research fellow under the supervision
of Dr. Yan.
Ultrasonic Wave Area
Device
Ultrasonic
Guided Wave
Laptop
PZT transducer
Table
Cubicle
Panel
Fig. 1: SurfingAttack leverages ultrasonic guided wave in the
table generated by an ultrasonic transducer concealed beneath
the table.
with voice assistants, such as Bixby, Siri, Google Assistant to
arrange appointments on the calendar or to start the morning
coffee brewing. While these new technologies significantly
improve the living quality, they also change the landscape of
cyber threats. Recent studies show that it is possible to exploit
the non-linearity in microphone to deliver inaudible commands
to the system via ultrasound signals [41], [42], [44], [52].
DolphinAttack [52] by Zhang et al. was among the first to
demonstrate inaudible attacks towards voice-enabled devices
by injecting ultrasound signals over the air, which can launch
from a distance of 5ft to the device. Recognizing the limitation
in the range of the attack, LipRead [42] extends the attack
range to 25ft by aggregating ultrasound signals from an array
of speakers, which requires line-of-sight. While these two
attacks demonstrate the feasibility of voice command injection
via inaudible ultrasound, they focus solely on over the air
transmission, which leads to several inherent limitations due
to the physical property of ultrasound wave propagation in air,
such as significant performance degradation when there is line-
of-sight obstruction. However, sound wave is fundamentally
the transfer of acoustic energy through a medium. It can
propagate wherever vibration is possible, such as water and
solid materials, in which the propagation characteristics are
different from air. Furthermore, the current literature [42],
[44], [52] focuses mostly on one-way interaction, i.e., they
inject commands to voice assistants without expecting any
feedback. However, voice-activated devices are designed to
enable multiple rounds of interactions. While the previous
literature has identified the new attack vector, its potential in
multi-round communication has received little attention.
In this work, we aim to understand the new threats enabled
by inaudible signal injection using ultrasound propagation in
Network and Distributed Systems Security (NDSS) Symposium 2020
23-26 February 2020, San Diego, CA, USA
ISBN 1-891562-61-4
https://dx.doi.org/10.14722/ndss.2020.24068
www.ndss-symposium.org
solid media, and the possibility of realizing multi-rounds of
hidden communication with AI-based voice assistants. Using
our proposed attack, SurfingAttack
1
, we found that it is possible
to deliver various inaudible voice commands in ultrasound to a
wide range of target devices from different manufacturers via
different solid media. Due to the unique properties of guided
wave propagation, SurfingAttack not only enables attacks from
a longer distance with a lower power requirement, but also
eliminates the need to be in the line-of-sight for inaudible
command injection attacks. By capitalizing on the capabil-
ity to control feedback mechanisms via the initial injected
command, SurfingAttack also enables inaudible multi-rounds
of interactions between the attacker and the target device
without alerting users in physical proximity. Fig. 1 illustrates
one of the application scenarios of SurfingAttack, where a
malicious device is hidden beneath the table to converse with
the target device on the top. By injecting voice commands
stealthily, attackers can instruct the voice assistants to leak
various secrets, such as an authentication code for money
transfer sent via an SMS message. The leaked secret can then
be picked up by a malicious device hidden away and relayed
back to the remote attacker. By leveraging unique guided wave
propagation properties in solid media, SurfingAttack presents
a new genre of inaudible attack on voice-activated systems.
While conceptually simple, there are several major chal-
lenges in realizing this attack: (a) how to design a hidden signal
generator that can penetrate materials effectively and inject the
inaudible commands without facing the victim’s device? (b)
How to engage in multiple rounds of conversations with the
victim’s device such that the voice response is unnoticeable to
humans while still being recorded by a tapping device?
For the first question, while the characteristics of sound
wave propagation in solid material is well studied for specific
application domains such as structure damage detection [39],
adapting the technique to deliver inaudible commands presents
unique challenges, such as wave mode selection, vertical
energy maximization, and velocity dispersion minimization.
Traditional ultrasonic speakers, as used in previous attacks,
are not suitable for exciting guided waves in table materi-
als due to their transducer structures. In order to adapt to
the solid medium, we utilize a special type of ultrasonic
transducer, i.e., piezoelectric (PZT) transducer, to generate
ultrasonic guided waves by inducing minor vibrations of the
solid materials. However, due to the unique characteristics
of ultrasound transmission in different solid materials, the
selection of different modes of guided wave can lead to
significant differences in the attack outcome, compared to
the over-the-air delivery of manipulated signals. To enable
SurfingAttack, we redesign a new modulation scheme that
considers wave dispersal patterns to achieve optimal inaudible
command delivery. SurfingAttack presents two unique features.
First, the attack is omni-directional, which works regardless
of the target’s orientation or physical environment where the
target resides. Second, the success of the attack is not impacted
by objects on a busy tabletop. To the best of our knowledge,
we are the first to deliver inaudible commands to a variety
of mobile devices through ultrasonic guided waves in a busy
environment. For the second question, to enable inaudible
1
The attack excites the guided waves that surf in the “ocean” of materials
and reach the surface to launch attacks, and thus is dubbed SurfingAttack.
multi-rounds of interactions, a tapping device is added along
with the ultrasound transducer to capture voice feedbacks from
the device. In order to minimize the impact of the feedback
on the environment, an injected command is used to tune
the output of the device to the lowest volume setting, such
that the feedback becomes difficult to notice by users, but
can still be captured by a sensitive tapping device. We have
conducted a series of experiments to understand the feasibility
and limitation of such low-profile feedback.
Leveraging the low attenuation of guided waves in solid
material and a place to hide the attack device, SurfingAttack
can enable a variety of new attacks including not only the
non-interactive attacks such as visiting a malicious website,
spying, injecting fake information, and denial of service by
turning on the airplane mode, but also interactive attacks that
would require multiple rounds of conversations with the target
device, such as unauthorized transfer of assets from the bank.
To demonstrate the practicality of SurfingAttack, we build a
prototype of the attack device using a commercial-off-the-shelf
PZT transducer, which costs around $5 per piece. Using our
prototype device, we conduct the following two attacks as a
demonstration:
(1) Hacking an SMS passcode. SMS-based two-factor
authentication has been widely adopted by almost all major
services [17], which often delivers one-time passwords over
SMS. A SurfingAttack adversary can activate the victim’s
device to read SMS messages in secret thereby extracting SMS
passcodes.
(2) Making fraudulent calls. A SurfingAttack adversary can
also take control of the owner’s phone to call arbitrary numbers
and conduct an interactive dialogue for phone fraud using the
synthetic voice of the victim.
We have tested SurfingAttack on 17 popular smartphones
and 4 representative types of tables. We successfully launch
SurfingAttack on 15 smartphones and 3 types of tables. A web-
site is set up (https://surfingattack.github.io/) to demonstrate
the attacks towards different phones under different scenarios,
and various new attacks such as selfie taking, SMS passcode
hacking, and fraudulent phone call attacks. With the growing
popularity of mobile voice commerce and voice payments [25],
we believe the demonstrated interactive hidden attack opens up
new attacker capabilities that the community should be aware
of. In summary, our contributions are as follows,
• We present, SurfingAttack, the first exploration of
attack leveraging unique characteristics of ultrasound
propagation in solid medium and non-linearity of
the microphone circuits to inject inaudible command
on voice assistants. We validate the effectiveness
of SurfingAttack on Google Assistant of 11 popular
smartphones, and Siri of 4 iPhones. We also show the
attack is resilient against verbal conversations.
• We evaluate SurfingAttack on 4 representative types
of table materials. We find that SurfingAttack is most
effective through 3 types of tables: aluminum/steel,
glass, and medium-density fiberboard (MDF). No-
tably, SurfingAttack can achieve long-range attack of
30ft distance through a metal table (the longest table
we can acquire is 30ft). We also validate the effective-
ness of SurfingAttack on aluminum and glass tables
2
with different thicknesses (up to 1.5 inch aluminum
and 3/8 inch glass).
• We further explore the possibility to pair command
injection with a hidden microphone to enable hidden
conversations between the attacker and the victim
voice assistant. We demonstrate several practical at-
tacks using the prototype we build, including hacking
an SMS passcode and making a ghost fraud phone call
without owners’ knowledge.
• We provide discussions on several potential defense
mechanisms, including using the high-frequency com-
ponents of guided waves as an indication of intrusion.
II. BACKGROUND AND THREAT MODEL
In this section, we introduce the background knowledge of
inaudible voice attack and physics of ultrasonic guided waves.
A. Inaudible Voice Attack
Audio capturing hardware in voice-controllable systems
generally includes a micro-electromechanical system (MEMS)
sensor to convert mechanical vibration to a digital signal, one
or more amplifiers, a low-pass filter (LPF), and an analog-to-
digital converter (ADC) to retrieve the sound in the physical
world. Inaudible voice attacks leverage the non-linearity of
the microphone circuits to inject inaudible commands to these
systems. The nonlinear response is due to the imperfection
of microphone and amplifier circuits [16], [27]. Let the input
sound signal be s(t), the output of microphone can be written
as:
s
out
(t) = A
1
s(t) + A
2
s
2
(t), (1)
where A
i
(i = 1, 2) is the gain of s
i
(t), while the higher
order terms are ignored as they are typically extremely weak.
The non-linearity term s
2
(t) produces harmonics and cross-
products. With carefully-crafted input signals based on the
baseband signals of voice commands, the microphone with
non-linearity can recover the baseband signals using the cross-
product term at the low frequency. Let the baseband voice
signal be v(t), the modulated input signal for launching attack
is designed as:
s(t) = (1 + v(t))cos(2πf
c
t), (2)
where f
c
is ultrasonic carrier frequency. After passing through
the microphone, the recorded signal by the microphone be-
comes:
r(t) = A
2
(1 + 2v(t) + v
2
(t))/2, (3)
since the high frequency components will be filtered out
by LPF. If the voice command component v(t) dominates
in the recorded signal, the voice controllable systems will
recognize the command. Previous work [52] demonstrated that
the nonlinear effect of MEMS microphone can be best incited
by ultrasonic frequencies between 20 kHz and 40 kHz.
B. Ultrasonic Guided Waves
The ultrasonic guided waves propagating in free solid-
material plates are known as Lamb waves, which have distinct
characteristics compared with ultrasonic waves in air. Assum-
ing the wave motion takes place in the x
1
x
3
plane, propagating
Symmetric modes Anti-symmetric modes
x
3
= h
x
3
=-h
x
3
= 0
Fig. 2: Symmetric and anti-symmetric Lamb wave modes
schematic in a free table plate. Symmetric Lamb wave is a
family of waves whose motion (the arrows in the figure) is
symmetrical with respect to the midplane of the plate (the
plane x
3
= 0), and anti-symmetric wave is a family of waves
whose motion is anti-symmetric with respect to the midplane.
along the x
1
direction, as illustrated in Fig. 2. Assume the
table plate is a stress-free plate with thickness of 2h (along
the x
3
direction), i.e., the mechanical component of stress at
the surfaces of the plate is zero. While Lamb waves, made
up of a superposition of longitudinal and transverse modes,
transmit in a thin plate, their propagation characteristics vary
with excitation and structural geometry of the plate. Every
Lamb waveform belongs to one of two modes: symmetric and
anti-symmetric [23], as presented in Fig. 2.
For a symmetric mode waveform at a given frequency,
the displacement fields u
1
(along x
1
) and u
3
(along x
3
) of
the Lamb wave in the complex-value representation are given
by [23]:
u
1
(x
1
, x
3
, t) = (ikA cos αx
3
+ βB cos βx
3
)e
i(kx
1
−ωt)
u
3
(x
1
, x
3
, t) = (−αA sin αx
3
− ikB sin βx
3
)e
i(kx
1
−ωt)
,
(4)
where A and B are given as an eigenvector of:
−2ikα sin αh
k
2
− β
2
sin βh
k
2
− β
2
cos βh −2ikβ cos βh
A
B
=
0
0
.
(5)
For an anti-symmetric mode waveform, the displacement fields
are:
u
1
(x
1
, x
3
, t) = (ikC sin αx
3
− βD sin βx
3
)e
i(kx
1
−ωt)
u
3
(x
1
, x
3
, t) = (αC cos αx
3
− ikD cos βx
3
)e
i(kx
1
−ωt)
,
(6)
where C and D are given by:
2ikα cos αh
k
2
− β
2
cos βh
k
2
− β
2
sin βh 2ikβ sin βh
C
D
=
0
0
.
(7)
Here, ω and k are the angular frequency and the wavenumber
of the Lamb wave, respectively. In addition,
α =
s
ω
C
L
2
− k
2
, β =
s
ω
C
T
2
− k
2
, (8)
where C
L
and C
T
are the longitudinal and transverse wave
speeds that can be derived from:
C
L
=
s
E (1 − v)
ρ (1 + v) (1 − 2v)
, C
T
=
s
E
2ρ (1 + v)
, (9)
3
Fig. 3: The phase-velocity dispersion curve of a glass table.
which depends on three material parameters: E is the Young’s
module, v is the Poisson ratio, and ρ is the density of plate
material.
The solutions of A, B, C, D in Eq. (5) and Eq. (7)
require the determinants of the two matrices to vanish, which
leads to the Rayleigh-Lamb equations (omitted for brevity) for
the plate. The Rayleigh-Lamb equations quantify the relation
between the angular frequency ω and the phase velocity C
p
of
the guided wave in the plate, where C
p
is the propagation speed
of the wave phase at a particular frequency within the guided
wave signals. C
p
relates to the wavenumber k: C
p
= ω/k.
According to Rayleigh-Lamb equations, Lamb waves exhibit
velocity dispersion; i.e., their propagation velocity depends on
the frequency (or wavelength) and material parameters.
The phase-velocity dispersion curve, which depicts the
correlation between the phase velocity and the frequency
thickness product (i.e., the product of wave’s angular frequency
and material thickness, ωh), is plotted in Fig. 3 for a glass
plate. We notice that with the increasing frequency thickness,
more propagation modes appear, i.e., more types of waves
with different transmission speeds will be propagating in the
material and interfering with each other. As a result, the center
frequency and guided wave mode of attack signals should
be carefully selected based on attack target information ex-
tracted from guided wave sensitivity studies. Different guided
wave modes introduce different wave propagation formats that
could significantly impact the effectiveness of attack signal
delivery, and the details of mode selection are presented in
Section III-A.
C. Threat Model
The attacker’s goal is to remotely converse with victim’s
voice controllable device to inject unauthorized voice com-
mands or to access sensitive information without victim’s
knowledge. We assume the victim is familiar with his/her
surrounding. This can be his/her own office or home, and
he/she can notice any physical alteration.
Physical Access. We assume that adversaries can place a
small attack device in the physical space of the device as long
as it is not visible to the user. We assume that an adversary
cannot physically touch the victim’s devices, alter the device
settings, or install malware apps. The activation commands
(“Hey Siri”, “OK Google”) of voice assistants are generally
voice fingerprinted, i.e., user verification is performed to
authenticate these commands. We assume the attacker can
synthesize the legitimate user’s voice signals using known
techniques [1], [33] to launch the attack when the target device
is voiceprinted.
No Owner Interaction. We assume that the target device
is placed on a medium that allows acoustic transmission,
such as a tabletop, and it is not being actively used by the
user. For smart home devices, the owner often interacts with
it less than a dozen times a day. For mobile phones and
tablets, it is also fairly common the owner is focusing on
other activities not related to the device, such as reading books,
having conversations with friends, working on a computer, etc.
Hidden Attack. One goal of the adversary is to attack
voice assistants without being detected. The adversary will
send the voice commands in ultrasonic frequencies that are
inaudible to humans, and at the same time, turn down the
volume of the device to the extent that it would be difficult
for the users to notice the voice responses from the assistant,
yet a hidden tapping device placed underneath the table can
record them.
Attack Equipment. We assume that adversaries possess
both the Piezoelectric (PZT) transducer designed for exciting
ultrasonic guided wave and commodity devices for generating
command signals. An ultrasonic signal source made of PZT
transducer is relatively small and can generally be concealed
and attached to a physical medium, such as the bottom of a
table.
III. KEY ELEMENTS OF SurfingAttack
There are three necessary conditions for the success of
SurfingAttack: (1) The ultrasonic wave in the table must be
able to reach the device microphone embodied in the device
enclosure. (2) Even when the microphone may not be in direct
contact with the transmission medium, the wave should still be
able to leverage the non-linearity of the device microphone on
the tabletop to launch the inaudible command injection attacks.
(3) The response from the victim device can be received by
the attacker via the planted device without raising suspicion
of the victim user. More specifically, the volume of victim’s
device can be tuned down such that user cannot notice it, yet
the response can be recorded by a tapping device beneath the
table.
A. Attack Wave Mode Selection and Generation
The first condition for the attack is the capability to deliver
inaudible ultrasound waves to the target device effectively.
Different from waves in air, the acoustic waves propagating in
solid materials have acoustic dispersion phenomenon, during
which a sound wave separates into its component frequencies
as it passes through the material. Lower dispersion indicates
a better concentration of acoustic energy. This implies that a
proper Lamb wave mode for SurfingAttack should feature (1)
low dispersion, (2) low attenuation, (3) easy excitability [46],
and (4) high attack signal reachability. To achieve the afore-
mentioned features, there are three key design decisions: the
signal waveform, Lamb wave mode, and the ultrasound signal
source.
First, guided wave signals can be generated via either
windowed modulation or pulse signals. It has been shown
that narrowband input signals are most effective in restricting
4
wave dispersal in large and thick plates [50]. As a result,
narrowband windowed modulation signals is used to carry the
attack command in SurfingAttack to minimize dispersion.
Second, different Lamb wave modes have different field
distributions throughout the whole plate [11], depending on
the different frequency-thickness and materials parameters as
shown in Fig. 3. Since the attack frequency range from 20
kHz to 40 kHz has the best performance in stimulating the
microphone’s non-linearity effect, we are limited to the lower-
order Lamb wave modes, i.e., A
0
or S
0
mode. In order to
succeed in the attacks, the Lamb wave should be able to spread
from a point of the table to the victim’s device on the tabletop
effectively. As a result, the generated Lamb waves need to
produce a high out-of-plane displacement
2
on the table surface.
As most of the displacement of the A
0
mode is out-of-plane,
while most of the displacement of the S
0
mode is in-plane
3
with lower frequency-thickness products. A
0
wave mode below
the cut-off frequencies of the higher order Lamb wave modes
is selected to create the ultrasound commands.
Lastly, we choose to use a circular piezoelectric disc
to generate the signal for its energy efficiency and omni-
directivity. It applies a vertical force towards the table surface,
resulting in a flexural wave propagating radially outwards and
thus enabling an omni-directional attack through the table.
The energy efficiency is important since the piezoelectric disc
that is hidden under the solid materials needs to produce
strong waves to reach extended distances with a minimal
amount of energy. The omni-directivity is crucial because
the attack should work regardless where the target’s location
and orientation are on the medium, i.e., wherever your phone
is placed on the table. The omni-directivity of the attack is
evaluated in Section VI-C. Furthermore, since objects on the
table surface could change frequently, we need to make sure
that the signal propagation still works regardless of whether
there are objects on the table. The corresponding evaluation is
presented in Section VI-G.
B. Triggering Non-linearity Effect via Solid Medium
While the non-linearity has been demonstrated for ultra-
sound wave that is directly delivered to the speaker via air, it is
unclear if it is feasible to trigger the same effect when acoustic
waves pass through the table materials to reach the external
enclosure of the phone. We conduct extensive experiments to
verify if the non-linearity effect of the voice capture hardware
of a smartphone placed on the tabletop can be triggered by the
ultrasonic guided waves that propagate in the table. The setup
for one of the initial experiments is shown in Fig. 4.
We use a low-cost radial mode vibration PZT disc [45]
(which only costs $5 per piece) with 22 mm diameter and 0.25
mm thickness to generate the ultrasonic guided wave. The disc
is adhered to the underside of an aluminum plate with 3 mm
thickness. The size of the PZT transducer is much smaller than
the ultrasound speaker used in existing attacks [42], [44], [52],
making the attacks more stealthy and economically accessible,
as shown in Fig. 4(b). We use a chirp signal from 50 Hz to
2
Out-of-plane displacement is defined as the displacement along the x
3
direction.
3
In-plane displacement is defined as the displacement along the x
1
direc-
tion.
SmartphoneMEMS microphone
Table
PZT transducer
(a) Schematic diagram
Smartphone
Waveform generator
PZT transducer
Aluminum plate
Back view
(b) Experimental setup
Fig. 4: An illustration of the experimental setup for investigat-
ing the feasibility of SurfingAttack.
(a) (b)
Fig. 5: Results of non-linearity test: (a) spectrogram of the
chirp baseband signal; (b) spectrogram of the recorded voice
signal by the smartphone when the signal frequency is 25.3
kHz.
5 kHz as the baseband signal. The baseband signal is then
imported to a Keysight 33500B series waveform generator and
modulated onto a carrier. The 9V output is then supplied to
the PZT transducer to excite Lamb waves. By analyzing the
recorded signal of a smartphone (i.e., Google Pixel), the non-
linearity of microphones could be evaluated.
Fig. 5 shows the spectrogram of the baseband signal
and the recorded signal when carrier frequency f
c
= 25.3
kHz. The ultrasonic guided wave propagates to the device
microphone and any resulted sound is recorded. The results
confirm the existence of the nonlinear response of the voice
capture hardware incited by ultrasonic guided waves. Fig. 5(b)
shows the recorded sound signal in the time-frequency domain,
in which the first harmonic component is almost identical to the
original signal displayed in Fig. 5(a). This result demonstrates
the feasibility of attacking voice controllable systems placed
on the tabletop through ultrasonic guided waves.
C. Unnoticeable Response
Unnoticeable response of a target phone is critical for
keeping the attack under the radar. Sound pressure level (SPL)
is used to quantify the sound pressure of a sound relative to a
reference pressure at the eardrums of our hearing or on the
diaphragms of the microphones. SPL is determined by the
corresponding audio voltage, while standard reference sound
pressure p
0
= 20 µ Pa ≡ 0 dB is the quietest sound a human
can perceive [49].
5
SPL depends on the distance between the area of mea-
surement and point-shaped sound sources in the free field. We
assume r
1
as the distance between the tapping device and the
sound source, r
2
as the distance between the user and the sound
source. L
1
and L
2
are SPLs at the tapping device and the user
end, the relationship of which follows the inverse distance law
as written below:
L
2
= L
1
−
20 · log
10
r
1
r
2
. (10)
Approximately, an SPL drop of 6 dB is expected by doubling
the difference of r
1
and r
2
. When SPL at the user end drops
below 0 dB, the voice response becomes essentially inaudible
to the user. Thus, it becomes feasible to conceal SurfingAttack
by adjusting the volume of the device via ultrasonic guide
wave and placing a hidden tapping device closer to the victim’s
device underneath the table. Note that the inverse distance
law is always an idealization because it assumes exactly equal
sound pressure as sound field propagates in all directions. If
there are reflective surfaces in the sound field, the reflected
sounds will be added to the direct sound, resulting in a higher
SPL at a field location than the inverse distance law predicts.
If there are barriers between the source and the point of
measurement, we may get a lower SPL.
To validate the feasibility, we evaluate the SPL of a Google
Pixel phone at different volumes, the results of which are
shown in Fig. 6. Here, we let the phone produce 1 kHz
sinusoidal tones with low volume levels, and an A-weighting
SPL meter is used to measure SPL at various distances. The
experiment is conducted in a quiet office (about 400 square
feet) with an average background noise of 40.5 dB. Although
the SPL stays above 0 dB, it decreases with distance, and
the signal is quickly overwhelmed by environmental noise
after propagating 50∼100 cm at volume level 1∼3. We also
deployed a microphone as a tapping device underneath the
table, which is proven capable of recording the weak voice
responses. The results show that it is feasible for the attacker
to adjust the volume low enough to make the voice responses
unnoticeable by the user from a moderate distance, while a
hidden tapping device can still capture the sound. To enhance
the sound capturing capability, we can deploy multiple tapping
devices at different positions under the table to precisely
capture the weak voice responses from the device speaker as
well. In an environment with larger background noise, we can
adjust volume even higher without alerting the owner. Lastly,
the attacker can turn off the screen to further enhance the
stealthiness of the threat. In Section V-D, we run extensive
experiments to corroborate the stealthiness of SurfingAttack
by measuring the responses of victim phones in different
environments.
IV. ATTACK DESIGN
To enable interactive hidden attacks, SurfingAttack gen-
erates well-crafted ultrasonic guided wave commands such
that they can propagate along the table to control the voice
assistants. The attack system is designed to initiate commands,
record voice responses, and interact with victim devices.
Without loss of generality, we will present our system design
details using Google Assistant as a case study, and the same
methodology applies to other voice assistants (Siri, Bixby).
Fig. 6: The SPL test results of the Google Pixel phone at
different volumes (Volume 0 represents background noise).
Signal Modulation
& Voice Recording
Signal
Processor
Voice Signals
& Control Commands
Internet
Voice Response
Interactive Voice Commands
& Dialogue Generation
Ultrasonic
Transducer
Tapping Device
Attack Device Package
Voice
Commands
TTS
Module
Speech
Synthesis
Voice
Recording
Controller
Fig. 7: SurfingAttack system architecture.
A. Attack Overview for SurfingAttack
Fig. 7 shows the system overview of the attack, where
the attacker planted the SurfingAttack device in the physical
environment of the victim, and the controller that supports the
major functionality is remotely connected off-site. Note that
the controller can be included in the SurfingAttack device as
well, however, that could change the computation requirement
and form factor of the SurfingAttack device. The SurfingAttack
device contains three main components: a signal processing
module, an ultrasonic transducer, and a tapping device, as
shown in Fig. 7. The primary function of the SurfingAttack
device is to enable the collection of voice device output and
the delivery of malicious commands via inaudible ultrasound.
The general workflow is as follows.
The voice commands or dialogues are generated using
the speech synthesis and text-to-speech (TTS) Module. The
controller produces the baseband signals v(t) of the voice
commands or dialogues, and then transmits them to the attack
device preferably through wireless, e.g., WiFi. The attack
device hidden beneath the table is used for ultrasonic signal
modulation and voice recording. The signal processor mod-
ulates the received baseband signal into the excitation signal
e(t) in Eq. (11) below. It is worth noting that: according to
the Nyquist theorem, the sampling rate of e(t) must be at
least twice the highest frequency of the signal to avoid signal
aliasing. The signal processor can be in the form of a portable
mobile phone with a relatively high sampling rate such as
Samsung Galaxy S6 Edge, with the addition of an amplifier
connected to the ultrasonic transducer. The transducer then
transforms the excitation signal into ultrasonic guided wave to
be propagated through the materials. Meanwhile, the tapping
device will record the responses, which are transferred back to
the controller in real time. Based on the responses, the attacker
can create the followup commands through controller. As such,
the interaction continues.
B. Ultrasonic Attack Signal Generation
Without direct control over the voice controllable system,
the attacker needs to carefully design inaudible voice com-
mands. In particular, SurfingAttack produces the modulated
6
signals of voice commands that can propagate in the table
to be received by the device’s microphone through mechanical
coupling.
Unlike the ultrasonic attack over the air, narrowband
window functions are used to modulate signals to reduce
wave dispersal, and the excitation signal must be preprocessed
before stimulating the guided Lamb waves. The signal pre-
processing ensures: (1) an appropriate frequency bandwidth
of the excitation signal in order to reduce signal distortion
due to dispersion; (2) a properly modulated signal to avoid
introducing audible sound.
Traditionally, guided wave testing uses a limited cycle
sinusoidal tone burst, which is often modulated by a Hann
window [46]. However, the Hann window eliminates high-
frequency signal characteristics. In order to preserve the simi-
larity between the recovered voice signal and the original sig-
nal, Tukey window (also known as the cosine-tapered window)
is used for modulation to form the excitation signal, described
as follows:
e (t) = (1 + m · v (t)) · w (t) · cos (2πf
c
t) , (11)
where m is the depth of the modulation, which can be selected
in [0.8, 1] based on empirical experimental results, and w(t)
represents the Tukey window:
w(t) =
1
2
1 + cos
2π
r
t −
r
2
,
1,
1
2
1 + cos
2π
r
t − 1 +
r
2
,
0 ≤ t <
r
2
r
2
≤ t < 1 −
r
2
1 −
r
2
≤ t ≤ 1
(12)
where r is a real number between 0 and 1. If r = 0, it
returns to a rectangular window; if r = 1, it becomes a Hann
window. In this study, we adjust r in the range of [0.1, 0.5] to
achieve the best attack performance on different target phones.
Tukey window offers a flat top window shape to control the
amplitude attenuation of time-series data, which protects v(t)
from distortion. The modulated signals will be used as attack
signals for the PZT transducer to excite ultrasound waves.
To activate the voice assistants, the baseband signal v(t)
will embed the wake words such as “OK Google” in front
of the attack commands. We use existing speech synthesis
techniques to generate the wake words of a specific voice [33],
and the attack commands can be simply generated using TTS
systems. However, in our experiments, we discover that after
the activation command wakes up the assistant, the device
creates a short vibration for haptic feedback to indicate the
assistant is ready. This vibration may negatively affect the
mechanical coupling, and thus reduce the attack success rate
of the subsequent attack commands. In response, we insert
a multi-seconds gap between the wake words and attack
commands to eliminate the vibration’s impact.
C. New Attacks Enabled by Interactive Hidden Attack
All mobile phones provide voice call service and Short
Message Service (SMS). Text messages or phone calls from
acquaintance are usually considered safe and trustworthy. With
the growing popularity of two-factor authentication, phone
verification has become one major means for identity authenti-
cation in a wide variety of web applications, including banks,
social networks, retail stores, email services, etc. Therefore, it
can be a serious threat if the attacker is able to unnoticeably
OK Google, Turn Volume to 3
Read my messages
You have one text message. It’s
from 347268, do you want to hear it?
Sure
It says …, do you want to reply,
repeat it and just that for now?
Cancel
Cancelled
(a) The SMS hacking attack
procedure.
OK Google, Turn Volume to 3
Call Sam with speakerphone
OK, calling Sam with speakerphone
Hi, Sam, I forgot the new access
code of the lab, can you tell me?
Sure, it is 2501.
OK, thanks.
You are welcome.
Sam
Sam
Fraud call using synthetic voice of Alice
Calling Sam
Sam Hi, Alice.
(b) The fraud call attack procedure.
Fig. 8: The procedure of SurfingAttack.
control the victim’s phone to read/reply/send text messages,
or make fraud calls to friends through a synthesized voice.
Here, we weaponize SurfingAttack to show its real-world threat
by demonstrating an SMS passcode hacking attack and a
fraud call attack (check out video demo in https://surfingattack.
github.io/).
SMS Passcode Hacking. Texting while driving has been
widely regarded as a dangerous activity for both the drivers
and the pedestrians. As a result, most virtual assistants have
offered features for listening and replying to text messages just
using voice commands, e.g., “read my messages” command
for Google Assistant or Siri, or “show me the most recent
message” command for Bixby. However, these features open
up opportunities for attackers as well. Moreover, the impact
of SMS hacking has been magnified due to its adoption as
the most universal and convenient two-factor authentication
technique. We describe the details of the SMS passcode
hacking attack below.
The attack procedure is displayed in Fig. 8(a). First, an
inaudible command “OK Google, turn the volume to three”
will activate the assistant and turn down the volume. With
such a low volume, the device’s responses become difficult
to notice by human ears in an office environment with a
moderate noise level. When a new message with the passcode
arrives, the attack device sends “read my messages”. Once
the victim device receives the command, it displays the most
recent message, state the sender of the message, and ask if
the owner wants to hear it or skip it. The tapping device (i.e.,
a commercial microphone) captures the message and sends it
to the attacker. In response, a subsequent inaudible command
“hear it” is then delivered to victim device, with which the
assistant will read the contents of the message. Finally, the
tapping device underneath the table can capture the sound and
send it to the attacker to extract the passcode. This process
allows the attacker to extract the passcode, when the device
is placed on the tabletop and the assistant reads the message
without alerting the owner.
Fraud Call. Phone call is one the most common ways
of communication methods nowadays. There has been a sig-
nificant increase in the number of phone scams in the past
7
few years, resulting in a billion dollars of financial losses
for consumers and businesses [38]. In general, it is common
for us to ignore unrecognized phone calls. Yet, when we
receive a phone call from an acquaintance or a contact, we
will subconsciously relax our vigilance. The advanced phone
scams rely on caller ID spoofing to deceive the victims into
believing that the call comes from a “trusted” caller, for which
effective defense mechanisms have been proposed [14]. Using
SurfingAttack, it is possible to place a fraud call attack via the
direct control of victim’s device, bypassing the caller authenti-
cation framework [8]. In this case study, we use SurfingAttack
to initiate a fraud call through victim’s device placed on the
tabletop without touching the device. SurfingAttack allows the
attackers to control the victim’s device to call someone and
conduct a multi-round conversation using the hidden ultrasonic
transducer and tapping device. Fig. 8(b) shows a fraud call
example in which the adversary controls Alice’s device to call
her friend Sam and deceive him into revealing the access code.
V. ATTACK EVALUATION
We validate SurfingAttack experimentally on 17 popular
mobile devices with intelligent voice assistants based on the
following three objectives: (a) examining the feasibility of
SurfingAttack; (b) quantifying the parameters in tuning a suc-
cessful attack; (c) measuring the attack command recognition
performance. This section describes the experimental setup and
results in detail.
A. Experimental Setup
Unless otherwise specified, all the experiments utilize the
default experiment equipment as shown in Fig. 4, which
consists of a waveform generator as the signal modulator, and
a PZT transducer to excite inaudible voice commands. All
experiments are conducted in an indoor lab environment with
an average background noise of around 40 dB SPL.
One key question is how generalizable the attack is, i.e.
how dependent the proposed attack is on the hardware and soft-
ware of the victim target device. To list a few considerations,
the materials of the phone body may impact the mechanical
coupling between the phone and the table; the software of the
voice assistants may implement defense modules to differenti-
ate between inaudible commands and human voice commands.
To systematically evaluate potential factors, we examine the
proposed SurfingAttack on 17 different types of phones with
the same experiment setup and attack equipment. We place
the victim devices 30 cm away from the PZT transducer on
two different types of plates: a rectangular frosted glass plate,
which is commonly used as tabletops in modern high-end
working desks, with dimensions of 24 × 30 × 1/16 inch, and
a steel metal plate with dimensions of 24 × 24 × 1/16 inch.
For each target device, we run three types of attacks:
recording activation, direct activation, and direct recognition.
For recording activation, we first allow the attack device to
send the activation command using ultrasonic guided waves
and record the sound. Then, the recording will be replayed
to attack the voice assistant. This attack tests if the recorded
sound has sufficient quality to perform the attack. For direct
activation, we send the activation command, i.e., “OK Google”
or “hey Siri”, directly via ultrasonic guided waves. For direct
Fig. 9: Mean amplitude of the demodulated chirps (from 5 Hz
to 5 kHz) baseband signal (1st harmonic) and 2nd harmonic
on Galaxy S7, with different f
c
.
recognition, we send inaudible voice commands directly to
control the device, such as: “call 123456”, “take a selfie”,
“read my messages”, after we manually activate the voice
assistant. The attacks are considered successful when the
assistant correctly executes the injected commands. We repeat
the experiment for each device when the device is either placed
facing up or facing down, and record the attack outcomes.
B. Feasibility Experiment Results
Table I summarizes the experimental results, which show
that SurfingAttack can successfully attack 15 out of 17 mobile
devices as listed, while the orientation of the devices (i.e.,
placed facing up or down) has negligible impacts on the attack
performance, which is likely due to the small form factor
of these mobile devices. Since different devices often have
different voice capture hardware (e.g. microphones, amplifiers,
filters), hardware layouts, and designs of the phone body, there
are always variations of digitized audios supplied to the voice
assistants. However, attack signal frequency f
c
is the dominant
factor that determines the attack’s feasibility. The average
amplitude of the demodulated chirp baseband signals, which is
used to evaluate the non-linearity property (see Section III-B),
is employed for selecting the optimal f
c
.
As an example, we measure the demodulated baseband
signal on a Samsung Galaxy S7 with various f
c
ranging from
24 kHz to 36 kHz, and compute the mean amplitude of the
demodulated chirp signal, as shown in Fig. 9. Ideally, the
demodulated baseband signal should have a high 1st harmonic
h
1
and a low 2nd harmonic h
2
. We select the optimal f
c
by considering the amplitude difference of the 1st harmonic
and 2nd harmonic of demodulated signals according to the
following formula:
f
opt
c
= argmax
f
c
((h
1
− h
2
) · h
1
).
For Samsung Galaxy S7, f
opt
c
= 25.8 kHz. We achieve
high non-linearity responses for the 15 devices which were
successfully attacked, but attain low non-linearity responses
for the other two devices. We find that all the 15 devices
are subject to all three types of attacks using different attack
frequencies. We also notice that both Xiaomi Mi 5 and
Huawei Mate 9 use the same Android 8.0, but Huawei Mate
9 successfully counters our attack. We believe that the phone
structure or the microphone used in Huawei Mate 9 suppressed
8
TABLE I: Experiment devices, systems, and results. The tested attacks include recording activation (record the ultrasonic
commands, and then replay it to the voice assistant), direct activation (activate the voice assistant), direct recognition (execute
voice commands). f
c
: attack signal frequency; m: modulation depth; r: cosine fraction of Tukey window; Mean Amplitude: the
average amplitude of the demodulated chirps at f
c
.
Manufacturer
Model OS/Ver. Assistants
Attacks
Best f
c
(kHz)
Best depth m &
cosine fraction r
Mean Amplitude
(dB)
Recording Activation Recognition
Google Pixel Android 10 Google X X X 28.2 m≥0.8, r=0.2∼0.5 -35.6
Google Pixel 2 Android 10 Google X X X 27.0 m≥0.8, r=0.2∼0.5 -35.0
Google Pixel 3 Android 10 Google X X X 27.0 m≥0.8, r=0.2∼0.5 -35.2
Moto G5 Android 7.0 Google X X X 27.0 m=1, r=0.3∼0.5 -35.2
Moto Z4 Android 9.0 Google X X X 28.2 m≥0.8, r=0.2∼0.5 -35.0
Samsung Galaxy S7 Android 7.0 Google X X X 25.8 m≥0.8, r=0.2∼0.5 -36.7
Samsung Galaxy S9 Android 9.0 Google X X X 26.5 m≥0.8, r=0.2∼0.5 -35.2
Xiaomi Mi 5 Android 8.0 Google X X X 28.3 m≥0.8, r=0.2∼0.5 -35.1
Xiaomi Mi 8 Android 9.0 Google X X X 25.6 m=1, r=0.3∼0.5 -35.6
Xiaomi Mi 8 Lite Android 9.0 Google X X X 25.5 m=1, r=0.3∼0.5 -35.3
Huawei Honor View 10 Android 9.0 Google X X X 27.7 m≥0.9, r=0.5 -35.0
Huawei Mate 9 Android 8.0 Google N/A N/A N/A 32.0 m≥0.8, r=0.1∼0.5 -75.6
Samsung Galaxy Note 10+ Android 10 Google N/A N/A N/A 26.0 m≥0.8, r=0.1∼0.5 -61.4
Apple
iPhone 5 iOS 10.0.03 Siri X X X 26.2 m≥0.8, r=0.5∼0.6 -37.2
iPhone 5s iOS 12.1.2 Siri X X X 27.1 m≥0.8, r=0.5∼0.6 -36.8
iPhone 6+ iOS 11 Siri X X X 26.0 m≥0.8, r=0.2∼0.5 -37.4
iPhone X iOS 12.4.1 Siri X X X 26.0 m≥0.8, r=0.2∼0.5 -37.3
ultrasonic signals. In the next section, we analyze the reason
why SurfingAttack fails.
The vibration of the table materials could also cause the
vibration of air around the table surface and be transformed
into ultrasonic waves in the air. Here, we design an experiment
to verify that the modulated ultrasonic wave is transmitted
through the solid medium to reach the microphone rather than
through the air. First, we use the in-air ultrasonic attack (i.e.,
DolphinAttack) to wake up the assistant of MI 5, which is
successful; and then, we use cotton and tape to block the
acoustic channel of the device’s microphone. We launch the
DolphinAttack for a second time and the attack fails, since
the ultrasound waves cannot enter the microphone. However,
even though the in-air channel towards device microphone is
blocked, we verify that SurfingAttack still succeeds.
C. Analysis of Failure Cases
In our experiments with 17 phones, we come across two
failure cases, including Huawei Mate 9, Samsung Galaxy Note
10+. Both phones have a curved back cover, and the Note 10+
also has a curved front screen. In order to trace the root cause
behind the failure, we install LineageOS 16.0 [37] on both
Xiaomi Mi 8 and Samsung Note 10+. With the same Android
OS, we eliminate the variation brought by different OSs. We
launch SurfingAttack towards these two phones equipped with
the same LineageOS, and the result shows that SurfingAttack
successfully attacks Xiaomi Mi 8, but still fails to attack
Samsung Note 10+, which indicates that the attack failure
cannot be attributed to the OS customization. Moreover, we
notice that the recorded sound of the ultrasound commands
from Samsung Note 10+ has a very weak strength, which is
likely caused by signal dampening over the body of the phone.
Therefore, our conclusion is that the failure of the attack is
most likely attributed to the structures and materials of the
phone body.
D. Stealthiness Experiment Results
Guided wave has an ultra-low magnitude (in the order of
microstrain or nanostrain). The vibration is relatively minor,
which is unlikely to be sensed by users even with a significant
increase in the transmission power. As a result, it is highly
unlikely for users to feel any vibrations during the attack. To
evaluate the stealthiness of our attack when the voice assistants
respond, we set up three experiments to measure the sound
levels of phones’ audible responses in different scenarios.
In the first experiment, we use a decibel meter to measure
the responses from Google assistant at different distances.
Specifically, we send voice command to Google assistant of
Google Pixel phone to read a long text message, and measure
the sound level of Google assistant’s responses at different
distances between the phone and the meter. Fig. 10(a) shows
the responses’ sound levels at phone’s volume-level 1∼3 in a
quiet lab environment. The dotted line represents the ambient
noise level at 43 dB, and the sound level below this line
would be difficult to recognize. The result shows that the
responses will be buried in ambient noise when the distance
goes beyond 50 cm. For the lowest volume setting (i.e., level
1), the responses will be hard to recognize with a distance
of around 25 cm
4
. Fig. 10(b) presents the sound levels in a
4
Generally, a phone placed on a table is at least 30 cm away from the
owner’s ear.
9
0 50 100 150 200
Distance (cm)
40
45
50
55
Level (dBA)
Vol=1
Vol=2
Vol=3
Env
(a) Google assistant’s response in a
quiet environment.
0 50 100 150 200
Distance (cm)
58
60
62
64
66
68
70
Level (dB)
Vol=1
Vol=2
Vol=3
Vol=4
Vol=5
Env
(b) Google assistant’s response in a
noisy environment.
0 10 20 30 40
Distance (cm)
41.6
41.8
42
42.2
42.4
42.6
42.8
Level (dB)
Vol=1
Vol=2
Vol=3
Vol=4
Vol=5
Env
(c) Incoming phone calls in a quiet en-
vironment.
0 10 20 30 40
Distance (cm)
58
60
62
64
66
Level (dB)
Vol=1
Vol=2
Vol=3
Vol=4
Vol=5
Env
(d) Incoming phone calls in a noisy
environment.
Fig. 10: The sound levels of responses at different distances in different environments.
0 10 20 30 40
Distance (cm)
0
0.2
0.4
0.6
0.8
1
Recognition rate
Vol=1
Vol=2
Vol=3
Vol=4
Vol=5
Fig. 11: Shazam’s music recognition rate with a victim phone
playing songs at different volume levels.
noisy environment at a McDonald’s restaurant with an average
noise level of 65 dB. The result shows that the responses
with volume-level 1∼5 will mostly be buried in ambient noise.
These experiments demonstrate the feasibility of SurfingAttack
especially in a noisy environment.
The second experiment evaluates the sound level in the
lowest volume setting when a callee responds to a fraud phone
call as described in Section IV-C. In this experiment, the phone
placed on the tabletop calls a callee, and the callee responds
with a long sentence at normal voice volume. Figs. 10(c)
and 10(d) present the sound level measurements in both quiet
and noisy environments. The results show that the callee’s
responses will be completely covered by the environmental
noise at volume level 1∼5.
To further show the feasibility of the attack, in our third ex-
periment, the victim Pixel phone plays songs from Youtube at
the lowest volume levels, and then we configure another phone
(iPhone 7+) to try to recognize the songs using the Shazam
app at different distances. The recognition rate performance
is depicted in Fig. 11, which shows that the recognition rate
drops to 0 with volume-level 1∼2 beyond 20 cm.
Finally, we ask 5 volunteers in the lab to act as victims of
SurfingAttack. Without checking their phones’ screens, none of
them is able to hear or feel the attack when the attack activates
their phones and interacts with them. Moreover, if their phones
are placed facing down, i.e., when the screens are invisible, it
becomes even more difficult to notice the attack. We record a
video to show the stealthiness and practicality of SurfingAttack
through the link https://youtu.be/zgr-oM2YJHs, in which we
can see there is not even a slight disturbance on a cup of water
TABLE II: The impact of background noises for activation and
recognition evaluated with Google Pixel.
Scene
Noises (dB)
“OK google”
“Read my
message”
Office 55-65 100% 100%
Cafe 65-75 100% 90%
Restaurant/Airport 75-85 100% 80%
during the attack.
VI. ENVIRONMENTAL IMPACT ANALYSIS
In this section, we evaluate the performance of
SurfingAttack in terms of the impacts of different background
noises, verbal conversations, directionality, attack distances,
table materials, table thicknesses, as well as the interlayers
and objects on the table. Unless specified otherwise, all the
experiments are conducted on both the frosted glass and steel
metal plate table tops.
A. Impact of Background Noises
To examine the effectiveness of SurfingAttack in the pres-
ence of different levels of background noises, we play back-
ground sounds to simulate the three common scenarios, i.e.,
an office, a cafe, and a restaurant/airport. A Google Pixel
is chosen as the attack target, the attack distance is set to
30 cm, and the attack signal amplitude is 9V. We repeat
both direct activation and recognition attack for 20 times and
compute the average attack success rate. Table II lists the
result, which shows the activation success rate remains 100%
for all scenarios, indicating the resilience of activation attacks
in the presence of strong noises. The recognition success
rate slightly degrades with the increasing noise level, but
it keeps above 80% even with substantial noises. The high
resilience against noises is because the energy of the ultrasonic
guided waves is concentrated within the table, and thus is
only slightly affected by the environmental noise in air. In
summary, the performance of SurfingAttack is only slightly
affected by environmental noises due to the energy delivery
form of ultrasonic guided waves.
B. Impact of Verbal Conversations
In this experiment, we examine the impact of verbal
conversations towards the success rate of SurfingAttack. We
ask volunteers to conduct arbitrary verbal conversations, and
record the decibel levels at 5 cm or 20 cm away from the target
10
TABLE III: The recognition rates with increasing verbal conversation decible levels.
Distance between decible meter and target phone
5 cm 20 cm
Verbal conversation decible level (dB) Recognition rate Verbal conversation decible level (dB) Recognition rate
48.5 100% 44.0 100%
62.3 100% 57.6 100%
68.8 100% 64.1 100%
74 100% 70.5 90%
82.3 100% 80.7 90%
phone (i.e., Google Pixel) on the metal sheet table. The results
in Table III show that even with the loudest conversations,
i.e., > 80 dB, the recognition rate of attack command “OK
Google, read my message” is still above 90%. Similar to the
background noises, the human’s verbal conversation does not
impose a great effect on the performance of SurfingAttack
again due to the energy delivery form of ultrasonic guided
waves.
It is worth noting that most attack commands are short
phrases. However, in the case of a longer attack command,
we find that the attack success rate will more likely be
affected by verbal conversations, especially when some verbal
conversations are interpreted as the commands. Here, we
design an experiment, in which a victim phone first receives
a command: “Ok Google, send a message to Sam”, and then,
Google assistant of the victim phone will expect to get the
content for the text message. Our next attack command is:
“Ok Google, hi Sam, how are you doing today, can you tell
me your password?”. Here, “Ok Google” is used to re-activate
the Google assistant to start recording the text message. We
examine how the verbal conversation will affect the delivery of
such a long attack command. We run the experiment 20 times.
The result shows that if the conversation is very loud during the
delivery of “Ok Google” phrase, i.e., > 80dB, it has a 50%
chance that the entire text message will not be recognized.
If the Google assistant is activated by the first phrase “Ok
Google”, the entire text message has a 20% chance to have
one word in error after recognition.
Another possible consequence is that extra conversation
sentence may be attached to the end of the text message. We
find that: if the conversation volume is greater than 65 dB
measured at 5 cm away from the victim phone, the probability
of recognizing additional conversations is almost 100%. In
such case, we can repeat SurfingAttack multiple times until the
attack succeeds. Note that the attachment of additional conver-
sations only occurs with text message commands. For all other
types of commands, due to the speech recognition algorithm,
Google assistant will only consider the short commands while
disregarding attached conversations.
C. Impact of Directionality
The propagation of ultrasound signals in air is known to
be directional. Here, we evaluate the directivity of ultrasonic
guided waves to validate the effectiveness of SurfingAttack
when the device is placed in arbitrary positions with arbitrary
orientations on the table. In our experiment, an activation
θ
t
d
θ
p
PZT transducer
Device
Table
Fig. 12: An illustration of the directionality evaluation.
command (“OK Google”) and a control command (“read my
message”) are used to test the directionality. Since microphone
sensitivity and the casing are different for each of the phones,
we select both Xiaomi Mi 5 and Google Pixel for this
experiment to measure the recognition rates of two commands.
The distance between the transducer and victim device is 30
cm. We evaluate both the impacts of the angle θ
p
(between
the axis of the mobile phone and the direct path) and the
angle θ
t
(between the direct path and the reference horizontal
axis) as shown in Fig. 12. The first experiment measures the
recognition rates at different θ
t
with a fixed θ
p
, and the second
experiment adapts θ
p
with a fixed θ
t
. For each command at
each position, we repeat it 20 times and calculate the average
recognition rate. The results of recognition rates for Google
Pixel keep at 100% regardless of its positions and orientations.
The results for Xiaomi Mi 5 are listed in Table IV. The
recognition rates for activation command remain as 100% for
all tests, while the recognition rates of the control command
also exceed 90% for all positions. The results demonstrate
the omni-directionality of ultrasonic guided waves, with which
SurfingAttack is able to attack the devices at arbitrary positions
and orientations on the table. In summary, SurfingAttack is
omni-directional, which can effectively target any devices
at arbitrary positions and orientations on the tabletop.
D. Impact of Attack Distances
In this section, we evaluate the recognition rates with vari-
ous distances between the attack device and victim device. We
repeatedly launch the activation command (“OK Google”) and
the control command (“read my message”) to a Google Pixel
at regular intervals and compute the probability of successful
attacks. Fig. 13 shows the recognition rate with increasing
distances on a glass table with a maximum length of 85 cm.
With a limited signal power (< 1.5 W ), the attack commands
in the form of ultrasonic guided wave can propagate over a
long distance without affecting the attack effectiveness. This
11
TABLE IV: The recognition rates at different θ
t
when θ
p
=0 (or
different θ
p
when θ
t
=0 with the results shown in parenthesis).
θ
t
(θ
p
) ”OK Google” ”Read my message”
0 100% (100%) 100% (100%)
45 100% (100%) 95% (100%)
90 100% (100%) 100% (90%)
135 100% (100%) 95% (95%)
180 100% (100%) 100% (100%)
225 100% (100%) 95% (100%)
270 100% (100%) 100% (95%)
315 100% (100%) 90% (100%)
20 30 40 50 60 70 80
Attack distance (cm)
0
0.2
0.4
0.6
0.8
1
1.2
Recognition rate
Ok Google
Read My Message
Fig. 13: The recognition rates with increasing distances be-
tween attack device and victim device.
can be attributed to the high energy conversion efficiency of
piezoelectric transducer and the low attenuation of ultrasonic
guided waves. As the activation command contains fewer
words than the control command, the recognition rates of the
activation command are slightly higher than that of the control
command.
Long Distance Attack. In order to execute the attack
experiment over a long distance, we set up a large table
using an Aluminum coil made of 6061 Aluminum alloy, with
the size of 6 inch × 30 feet. We use Google Pixel as our
attack target, with the attack message of “OK Google, read my
message.” To achieve a long distance attack, we amplify the
original signal using a commercial GWBP-AMP-X75 power
amplifier with maximum output power of 1.5W and maximum
output voltage of 30V . The result shows that SurfingAttack
successfully attacked Google Pixel phone placed at the furthest
end of this large table with 1.5W attack power and 28.8 KHz
attack frequency f
c
, with 100% of attack success rate (i.e.,
recognition rate). We believe that the attack distance can extend
even further than 30 feet. However, at the time of writing, the
longest Aluminum coil we were able to purchase was 30 feet.
In addition, we verify that the attack success rate stays above
80% when the attack power is reduced to 0.75W (with 15V
attack signal amplitude). In comparison, the furthest inaudible
attack distance over the air using an ultrasonic speaker array
is up to 30 feet using the attack power of 6W [42]. With 10%
of the attack power, SurfingAttack remains at least as effective
over a potentially longer distance on a large Aluminum table.
0 0.2 0.4 0.6 0.8
Attack power (Watt)
0
100
200
300
400
Achievable attack distance (cm)
Reaching 30 feet (900cm+)
Fig. 14: Relationship between attack power and distance.
Relationship Between Power and Distance. Given the
large Aluminum table, we further evaluate the relationship be-
tween the attack power and attack distance using Google Pixel
phone as a target. We repeat each attack 20 times, and record
the recognition rates across different distances. Since the attack
recognition rate does not have to be 100% for a successful
attack, we define the attack distance as the maximum distance
across which the SurfingAttack’s recognition rate exceeds 50%.
The result is presented in Fig. 14, which shows that the attack
distance has a positive correlation with the attack power. When
the attack power exceeds 0.8W (with attack signal amplitude
of 15V ), the attack distance reaches 30 feet, the maximum
length of our Aluminum table. It is worth noting that: even
with the highest power, the user will not be able to sense
any vibrations due to the attack signals’ energy delivery form,
as discussed in Section V-D. In summary, SurfingAttack can
effectively attack the voice assistants placed far away from
the attack device with high attack success rate.
E. Influences of Table Materials
The performance of SurfingAttack is heavily dependent on
both the materials and thicknesses of the tables which deliver
the ultrasonic attack signals to the voice assistants. The mate-
rials or thicknesses of the tables influence the characteristics
of the guided wave generation, propagation, and mechanical
coupling with the device. We provide a thorough study of
such impacts via both theoretical analysis and experimental
validation on four different types of tables.
Material Influence Analysis. Four most common table
materials, i.e., glass, metal, one type of wood: medium-
density fiberboard (MDF), and one type of plastic: high-
density polyethylene (HDPE), are selected for impact analysis.
Propagation of Lamb waves depends on the density and the
elastic material properties of the medium, the test frequency,
and material thickness, which are listed in Table X in Ap-
pendix. Fig. 15 displays the phase-velocity dispersion curves
of A0 mode in different tables at 20∼40 kHz attack frequency
range. Different dispersion affects the demodulated commands
since signals at different frequencies propagate in the table at
different speeds. Such effect becomes more significant in a
long-range attack.
In addition, given a PZT transducer, the excitation ampli-
tude of the guided wave also depends the thickness. Victor
[18] has proposed a theoretical model to compute the Lamb
12
20 25 30 35 40
Frequency (kHz)
0
500
1000
1500
2000
Phase speed (m/s)
5mm MDF
5mm Glass
2.54mm Glass
3mm AL
Fig. 15: The phase-velocity dispersion curves of A0 mode in
different tables at the attack frequency range.
0 5 10 15 20 25 30
Thickness(mm)
0
0.2
0.4
0.6
0.8
1
Normalized wave amplitude
Glass
MDF
Fig. 16: Predicted A
0
mode Lamb wave amplitudes by a 22
mm diameter PZT transducer w.r.t. the thickness of table.
wave response under PZT excitation. The strain amplitude of
the anti-symmetric mode A
i
(i=0,1,2,...) Lamb wave can be
calculated as follows [18]:
|ε
i
| =
aτ
0
µ
sin (k
i
a)
N
A
(k
i
)
D
0
A
(k
i
)
, (13)
where
N
A
(k) = kβ
k
2
+ β
2
sin (αh) sin (βh) ,
D
A
(k) =
k
2
− β
2
2
sin (αh) cos (βh)
+4k
2
αβ cos (αh) sin (βh) .
(14)
Here, a is the radius of the transducer, τ
0
is the surface
stress amplitude by the transducer, µ is the Lame’s second
constant, α and β are in Eq. (8), N
A
and D
A
can be found
in Eq. (14). The equation D
A
= 0 is the Rayleigh-Lamb
characteristic equation for anti-symmetric modes, and k
i
are
the simple roots of the dispersion equation. SurfingAttack’s
attack frequency range is limited to 20∼40 kHz due to the non-
linearity property of the victim’s microphone. Fig. 16 shows
the predicted maximum amplitude by a 22 mm diameter PZT
in the attack bandwidth w.r.t. the thickness of both glass and
MDF tables. As the thickness increases, the excitation of the
guided wave requires more energy since the energy will be
dispersed across the entire table thickness. However, only the
wave on the tabletop can reach the victim device, thus the
attack performance gets worse in a thicker table given a certain
attack power. For example, when the MDF/glass thickness
increases from 3 mm to 2 cm, the wave amplitude will be
reduced by 14 dB. We perform experiments to evaluate the
impact of table thickness in Section VI-F below.
TABLE V: Energy transmission coefficients of different mate-
rial pairs.
Device
Table
Aluminum Glass MDF
Ceramic 0.59-0.78 0.53-0.73 0.11-0.17
Aluminum 0.89-1 0.84-0.99 0.23-0.41
Glass 0.97-0.99 0.94-1 0.31-0.46
Influence of Table Materials Towards Reception Perfor-
mance of Victim’s Microphone. The reception performance
of the victim’s microphone depends on the table materials, the
difference of which brings in different mechanical coupling
over the boundaries. As the interaction of ultrasonic guided
waves with device boundaries is complicated, we conduct a
qualitative analysis based on acoustic impedance. Ultrasonic
waves are reflected at boundaries where there is a difference
in acoustic impedances (Z) of the materials on each side of the
boundary. This difference in Z is commonly referred to as the
impedance mismatch. The greater the impedance mismatch,
the higher the percentage of energy that will be reflected at
the interface or boundary between one medium and another.
As for SurfingAttack, the more energy that can be delivered
to the device, the greater attack success rate will be observed.
We can calculate the transmitted incident wave intensity based
on the fact that particle velocity and local particle pressures
must be continuous across the boundary. When the acoustic
impedances of the materials on both sides of the boundary
are known, the fraction of the incident wave intensity that is
transmitted through the boundary can be calculated as [5]:
T =
4Z
t
Z
d
(Z
t
+ Z
d
)
2
. (15)
This value is known as transmission coefficient, where Z
t
and Z
d
are the acoustic impedances of table and device,
respectively. The acoustic impedance depends on the density
and speed of sound, as shown in Table X. During the attack, the
transverse wave component in the table is the prime incident
wave, and both longitudinal and transverse wave components
can propagate into the device. The transmission coefficients
of different material pairs for device and table are listed in
Table V. Here, we consider three table materials, and three
device body materials including ceramic, metal, and glass.
Generally, the best energy delivery can be achieved when the
table material is the same as the device body material, which is
the reason why ultrasonic guided waves transmitting in MDF
tables result in lower energy delivery performance.
Evaluation Experiments. We run experiments to evaluate
the maximum attack distances on different tables with a limited
attack power. The results are shown in Table VI. Five tables are
evaluated in this study: an Aluminum metal tabletop with 0.3
mm thickness, a steel metal tabletop with 0.8 mm thickness, a
glass tabletop with 2.54 mm thickness, an MDF tabletop with
5 mm thickness and, an HDPE tabletop with 5 mm thickness.
With a limited attack power, the attacks fail for the HDPE
table, because of both the small acoustic impedance as shown
in Table X and rough matte surface of the tabletop. The results
for other tabletops are shown in Table VI. The experimental
results with different devices validate our theoretical analysis.
13
TABLE VI: Maximum attack distance on different tables (attack power is less than 1.5 W ). The width of Aluminum metal table
is 910 cm, the width of metal table is 95 cm, and the width of glass table is 85 cm.
Device
Max attack distance (cm)
Aluminum Metal Sheet (0.3 mm) Steel Metal Sheet (0.8 mm) Glass (2.54 mm) MDF (5 mm)
Activation Recognition Activation Recognition Activation Recognition Activation Recognition
Xiaomi Mi 5 910+ 910+ 95+ 95+ 85+ 85+ 50 47
Google Pixel 910+ 910+ 95+ 95+ 85+ 85+ 45 42
Samsung Galaxy S7 910+ 910+ 95+ 95+ 85+ 85+ 48 N/A
With the larger thickness and greater impedance mismatch,
the attack distances in the MDF table are much smaller than
the other tables. However, the attack range on an MDF table
can be increased by using a more powerful amplifier (in this
study, the amplitude of output is limited to 30V by the power
amplifier). For both metal and glass tables, a successful attack
can be executed with a much larger table as evidenced in
Section VI-D. In summary, SurfingAttack can effectively attack
the devices placed on tables with different table materials,
especially metal and glass materials. The attack performance
improves when the table material matches the device’s body
material.
F. Influences of Table Thickness
The surface materials of different tables may have different
thicknesses. In this experiment, we evaluate the influences
of different table thicknesses. We purchase different thick-
nesses of Aluminum sheets and glass sheets from McMaster-
Carr [34]. Specifically, Table VII lists the recognition rates
across different thicknesses of metal and glass tables for the
attack message “OK Google, read my message”. We present
the results with both 9V and 30V attack signal amplitudes. The
results with 9V attack signals show that the recognition rate of
the attack message degrades with increasing material thickness,
which matches with our theoretical analysis. With Aluminum
sheet, we find that the recognition rate starts dropping when the
thickness increases to 1/4 inch. With 1/2 inch thick Aluminum
sheets, the attack becomes unsuccessful: the recorded sound
shows a significant distortion of voice signals due to the
propagation complexity brought by the thick metal sheets.
However, with 30V attack signals, SurfingAttack succeeds with
100% rate even with the thickest 1.5 inch metal sheet.
In the glass material, we notice the same trend of de-
grading attack performance with increasing thickness under
9V attack signals. With 3/8 thick glasses, the recognition rate
of attack commands drops to 0%. However, with 30V attack
signals, SurfingAttack succeeds even with the thickest glass
table. Moreover, we have an interesting observation that, with
different thicknesses of materials, the best attack frequencies
f
c
are different. This can be attributed to the phase-velocity
dispersion of guided waves: as shown in Fig. 3, the propagation
of the guided wave becomes different with varied signal
frequencies and thicknesses, which leads to different attack
performance.
G. Impact of Interlayers and Objects on the Table
In a realistic scenario, the device may not be in direct
contact with the tabletop. There may exist one or more inter-
mediate layers between them, such as documents, newspapers,
tablecloths, and mobile phone cases, etc. When the incident
waves cross a layer sandwiched between two media, wave re-
flection and wave propagation through an elastic layer depend
on the frequency and interlayer thickness, which are associated
with resonances in the layer [7]. In this study, we place Google
Pixel 25 cm away from the PZT transducer on the glass
table with various interlayers, and measure the attack success
rate with the command “OK Google, read my message”, the
results of which are shown in Table VIII. For each layer, we
repeat the command 20 times and calculate the average success
rate. SurfingAttack exhibits excellent performance with most
interlayers. However, the attack fails to penetrate the interlayer
made of a Peva front and non-woven backing tablecloth, since
the impedance mismatch is intensely increased by the non-
woven layer. In another experiment, we place arbitrary objects
on the table, the recognition performance of SurfingAttack on
different devices is unaffected.
Phone cases are popular accessories for hardware protec-
tion. We run an additional experiment to evaluate the impact of
phone cases (mostly made of silicone rubber) for four different
phones. The results in Table IX show that the recognition rates
are only slightly affected by the phone cases. However, we note
that the performance degradation could be more significant,
if thicker phone cases made of uncommon materials such as
wood are used.
In summary, SurfingAttack can successfully attack devices
on the tables covered with most types of interlayers except for
some special types of tablecloths. Moreover, objects on the
table do not affect SurfingAttack.
VII. COUNTERMEASURE
In this section, we discuss the defense strategies to defend
against SurfingAttack.
Hardware Layout Enhancement. One prerequisite for the
success of the attack is that the ultrasonic voice commands can
propagate along the device body to the microphones. Thus,
the layout of microphone could be enhanced and redesigned
to damp or suppress any acoustic vibration whose frequencies
are in the ultrasound range.
Interlayer-based Defense. One effective but simple de-
fense mechanism of SurfingAttack is to place the device on a
soft woven fabric or a multilayers term (the peva & non-woven
two layers tablecloth in Table VIII) to increase the impedance
mismatch.
Software-based Defense. An ideal software-based defense
approach should identify and reject received voice commands
14
TABLE VII: The recognition rates with different table thicknesses and attack signals for Aluminum metal and glass.
Aluminum Metal Glass
Thickness (inch)
Recognition rate
(9V attack signal)
Recognition rate
(30V attack signal)
Thickness (inch)
Recognition rate
(9V attack signal)
Recognition rate
(30V attack signal)
1/16 100% 100% 1/16 100% 100%
3/16 100% 100% 1/8 100% 100%
1/4 70% 100% 3/16 50% 90%
1/2 0% 100% 1/4 10% 90%
1 or 1.5 0% 100% 3/8 0% 100%
TABLE VIII: The recognition rates with different interlayers.
Interlayer Thickness (mm)
Recognition rate
“OK google, read my message”
3 sheets of papers 0.3 75%
Hard plastic phone case 1.4 90%
Polyester tablecloth 0.5 95%
Vinyl table protector 0.3 90%
Two layers (peva & non-woven) tablecloth 0.7 0%
TABLE IX: The recognition rates of phones with and without
cases.
Device
Aluminum Metal Glass
w/o case with case w/o case with case
Xiaomi Mi 8 100% 90% 100% 100%
Huawei Honor View 10 100% 100% 100% 90%
Google Pixel 100% 100% 100% 100%
Moto Z4 100% 100% 100% 100%
0 0.5 1 1.5 2
Frequency (kHz)
0
0.02
0.04
Amplitude
Original
0 0.5 1 1.5 2
Frequency (kHz)
0
0.02
0.04
Amplitude
Recovered
Fig. 17: Frequency responses of original (top) and recovered
(bottom) voice signals after attack.
that are not the genuine voices by analyzing the unique features
of attack signals which are distinctive from the genuine ones.
Prior studies [52] have shown the difference between the
recovered ultrasound attack signal and the original signal
in the frequency ranging from 500 to 1,000 Hz. However,
with SurfingAttack, there is no significant difference in that
frequency range between the genuine signal and ultrasonic
signal as shown in Fig. 17.
Nevertheless, we discover a notable difference between
(a) (b)
Fig. 18: Time plots and spectrograms for: (a) normalized
original voice; (b) normalized recorded signal after attack.
the recovered attack signal and the baseband signal at the
high frequency ranging from 5 kHz to 20 kHz, as shown in
Fig. 18. The original signal is produced by the Google TTS
engine with the carrier frequency for modulation as 25.7 kHz.
The recovered attack signal of SurfingAttack suppresses the
frequency components from 5 kHz to 8 kHz and produces
a new frequency component from 10 kHz to 20 kHz mainly
attributed to the complicated nonlinear response. Since human
voice mainly occupies low frequencies, most feature recogni-
tion algorithms for speech recognition (such as Mel-frequency
cepstral coefficients or MFCCs) mostly focus on low frequency
features, while high frequency features take less weight. As a
result, even if the high frequency components of the signal are
distorted, the attack still sabotages many devices.
With such observation, we propose an attack index based
on the frequency response R(f) of the received signal as
defined below:
AttackIndex = log
R
f
2
f
1
R(f)df
R
f
1
0
R(f)df
. (16)
Here, f
1
is 10 kHz, f
2
is 20 kHz. To validate the feasibility
15
0 10 20 30 40 50 60
Test #
-8
-6
-4
-2
0
2
Attack Index
Original
Pixel Recovered
S7 Recovered
Fig. 19: Attack indices of original signals and recovered signals
(from ultrasound commands).
of identifying SurfingAttack, we generate 54 attacks with
different attack parameters (i.e., frequencies, table materials,
distances, baseband signals, devices). The attack indices of
both original signals and recovered signals are presented in
Fig. 19. By monitoring the attack index, we can effectively
detect SurfingAttack, since there are almost no such high
frequency components in the human voice. As such, if the
attack index is higher than a pre-set threshold, the received
voice signal would be classified as an attack. It is worth noting
that if the cut-off frequency of the audio low-pass filter in
the device is lower than 10 kHz (e.g., Xiaomi Mi 5), this
defense strategy may become invalid since the attack index
cannot be computed without the information of high frequency
components.
VIII. DISCUSSION
Attacking Standing Voice Assistants. Amazon Echo and
Google Home are standing voice assistants with microphones
distributed across the cylinder. The current SurfingAttack can-
not reach these microphones. We believe this is due to the
significant power loss during the power transition across the
boundary of the table material and speaker material, as well
as the devices’ internal construction in terms of the relative
position of the microphones. With a better amplifier, it could
be possible to excite high-power ultrasound signals to reach
the standing speakers’ microphones.
Short Attack Distance on MDF Table. As shown in
Table VI, the attack distance of MDF was significantly shorter
than that of other table materials, which we believe can also
be improved by increasing signal power. With 1.5 W attack
power, SurfingAttack reaches the maximum attack distance of
50 cm on MDF material. To improve the effectiveness of
SurfingAttack, the adversary can attach multiple transducers
distributed across the table, which can alleviate the short attack
distance limitation on MDF tables.
Compact and Portable Attack System. In the demon-
stration of SurfingAttack, we use a waveform generator for
signal generation, and the voice commands are uploaded to
the memory before every experiment. To build a compact and
portable attack system, we could use mobile devices to produce
modulated attack signals. Although most mobile devices can
only transmit a modulated narrow-band signal with the carrier
frequency of at most 24 kHz due to their sampling rate limit of
48 kHz, Samsung Galaxy S6 and Samsung Galaxy S7 support
96 kHz sampling rate, which can be used to generate attack
signals. A portable attack system using these mobile devices
as shown in Fig. 7 will be developed in future work.
Limitation of Interactive Attacks in Capturing Victim
Feedback. In order to avoid alerting the users, SurfingAttack
adjusts the victim phone’s volume to the lowest settings, which
also increases the difficulty of capturing the victim’s feedback
by a hidden tapping device. In the presence of significant envi-
ronmental noise, it could be challenging to capture and recover
the weak feedback. However, a highly-sensitive tapping device
such as a parabolic microphone placed underneath the table
can be used to improve the feedback capture efficiency. We
can also apply signal processing techniques to separate out
the feedback from the noise.
Attack Specific Devices or Multiple Devices Simulta-
neously. Table I shows that different best attack frequencies
f
c
work for different mobile devices, which provides an
opportunity to attack a specific device in the presence of
multiple devices. For example, Xiaomi Mi 8 or Mi 8 Lite
is subject to 25.5 KHz attack frequency, which is lower
than most of the other devices. An adversary could leverage
such a parameter difference for targeted attack. Meanwhile,
SurfingAttack could also attack multiple devices with similar
parameter settings simultaneously using non-interactive attack
commands, and we demonstrate such a multi-device attack on
the website. However, if some mobile devices have customized
wake-up words, SurfingAttack will not be able to activate them
simultaneously, but it offers another opportunity for launching
targeted attack when the attacker learns the specific wake-up
words.
Limitation of Voice Unlock. Unlocking the phone with
Google Assistant is as easy as speaking “OK Google”. How-
ever, we discover that after the Google Assistant upgrade in
March 2019, Google replaces the “voice unlock” with the “lock
screen personal results” function [21]. Equipped with this new
feature, the user can control what types of information the
Assistant will speak or show when the phone is locked. If the
victim happens to turn off lock screen personal results, the
attacker will need to unlock the device with fingers to hear
Assistant’s responses that include personal information. If the
victim turns on this feature (which is often the case), the attack
still succeeds. Therefore, locking the device and turning off
lock screen personal results function could be one solution to
defend against SurfingAttack. Note that only the pattern, PIN
and password screen lock can counter SurfingAttack, while the
swipe screen lock could not. Moreover, the recent Android and
iOS systems allow users to disable the voice assistants on lock
screen, in which case SurfingAttack can be prevented as long
as users lock their device every time they put it down.
IX. RELATED WORK
Voice-based Attacks. With the rapidly growing popularity
and functionality of voice-driven IoT devices, voice-based
attacks have become a non-negligible security risk. Gong et
al. investigate and classify voice-based attacks [20] into four
major categories: basic voice replay attacks [12], [29], [36],
operating system level attacks [3], [15], [26], [53], machine
learning level attacks [2], [9], [10], [13], [19], [43], [48], [51],
and hardware level attacks [28], [52]. A machine learning level
attack uses adversarial audio commands to attack automatic
speech recognition (ASR) systems. The commands are intel-
ligible to ASR systems, but cannot be easily comprehended
by humans. A hardware level attack replays a synthetic non-
speech analog signal instead of human voice. For example,
16
Kasmi et al. [28] introduce a voice command injection attack
against smartphones by applying intentional electromagnetic
interference on headphone cables to emit a carefully designed
inaudible AM-modulated signal. Our work can be considered
as one special type of hardware level attack.
Sonic or Ultrasonic-based Attacks. Recently, researchers
have proposed sonic or ultrasonic wave based attacks. Bolton
et al. [6] show that the crafted ultrasonic tones could corrupt
the hard drives and operating systems, causing spontaneously
reboot. Trippel et al. [47] gain control of the outputs of
MEMS accelerometers by leveraging the circuit imperfections
with resonant acoustic injections. A number of recent research
studies have focused on attacking the voice controlled system
using ultrasonic wave [41], [42], [44], [52]. These attacks
employ the ultrasound in air for delivering voice commands.
However, due to the directivity of ultrasound and unmaskable
size of the ultrasound speaker, the attack device can be easily
exposed. In contrast, we show that it is possible to use
ultrasonic guided waves to inject inaudible commands through
solid materials with a hidden attack device, causing a wide
variety of more serious security and privacy issues.
Guided Wave Technology. Recently, Roy et al. [40]
propose Ripple II, which is a surface communication scheme
through vibration. Ripple II excites acoustic vibrations to
microphone for linear response, while SurfingAttack excites
ultrasonic guided waves to microphone for nonlinear response.
The sources (vibration motor versus ultrasound sensor), the
principles of acoustic responses, and the purpose are different,
which lead to completely different signal designs. Moreover,
similar physical vibrations on a solid surface have been used
for user authentication [31], as well as touch location and
object identification [30]. These physical vibrations induced
by vibration motors have different characteristics, compared
with the insensible ultra-minor vibration of ultrasonic guided
waves used in this paper.
Ultrasonic guided waves have been proven useful for both
Nondestructive Testing (NDT) [39] for materials and Struc-
tural Health Monitoring (SHM) [46] for structures. Guided
wave testing (GWT) employs acoustic waves that propagate
along an elongated structure while guided by its bound-
aries [39]. This allows the waves to travel a long distance
with minor energy losses. GWT is widely used to inspect
and screen engineering structures [35], particularly for the
inspection of metallic pipelines around the world [4]. There
are also applications for inspecting rail tracks [11], rods [24],
and metal/composite plate structures [22], [32]. To the best
of our knowledge, this research is the first to present a novel
attack towards voice assistants using guided wave technology.
X. CONCLUSION
In this paper, we explore the feasibility of launching inaudi-
ble ultrasonic attack leveraging solid material as a transmission
medium. Compare to the previous studies on over-the-air
transmission, our proposed attack SurfingAttack can conceal
itself within/beneath the materials, offering new avenues to
launch inaudible attack in a previously unavailable setting.
Leveraging the energy delivery form of ultrasonic guided wave,
SurfingAttack proves to be an effective and economic attack,
successfully attacking devices across a long distance through a
30ft long table with only 0.75W attack signal power. Extensive
experiments were conducted to explore the extent of this
newly discovered threat as well as its limitation. Furthermore,
recognizing that voice controllable devices are designed to
enable conversations between human and computer, we further
extend our attack to listen to the voice responses with minimal
volume, enabling conversations between the adversary and the
voice controllable device. Using SurfingAttack, we demon-
strated potential attacks that will allow an adversary to hack
the SMS passcode or make a fraud phone call. We also provide
discussions and several defenses to mitigate SurfingAttack.
ACKNOWLEDGEMENT
The authors are grateful to the anonymous reviewers for
their constructive comments and suggestions. This work is sup-
ported in part by the National Science Foundation grants CNS-
1950171, CNS-1949753, CNS-1916926, and CNS-1837519.
REFERENCES
[1] “Modulate.ai,” https://modulate.ai/, 2019, accessed: 2019-09-07.
[2] H. Abdullah, W. Garcia, C. Peeters, P. Traynor, K. R. Butler, and
J. Wilson, “Practical hidden voice attacks against speech and speaker
recognition systems,” in Network and Distributed Systems Security
(NDSS) Symposium, 2019.
[3] E. Alepis and C. Patsakis, “Monkey says, monkey does: security and
privacy on voice assistants,” IEEE Access, vol. 5, pp. 17 841–17 851,
2017.
[4] D. Alleyne, B. Pavlakovic, M. Lowe, and P. Cawley, “Rapid, long range
inspection of chemical plant pipework using guided waves,” in AIP
conference proceedings, vol. 557, no. 1, 2001, pp. 180–187.
[5] D. T. Blackstock, “Fundamentals of physical acoustics,” 2001.
[6] C. Bolton, S. Rampazzi, C. Li, A. Kwong, W. Xu, and K. Fu, “Blue
note: How intentional acoustic interference damages availability and
integrity in hard disk drives and operating systems,” in 2018 IEEE
Symposium on Security and Privacy (SP), 2018, pp. 1048–1062.
[7] L. Brekhovskikh, Waves in layered media, 2012, vol. 16.
[8] callauthenticate, “Combating spoofed robocalls with caller id authenti-
cation,” https://www.fcc.gov/call-authentication, 2019, accessed: 2019-
09-07.
[9] N. Carlini, P. Mishra, T. Vaidya, Y. Zhang, M. Sherr, C. Shields,
D. Wagner, and W. Zhou, “Hidden voice commands,” in 25th USENIX
Security Symposium (USENIX Security 16), 2016, pp. 513–530.
[10] N. Carlini and D. Wagner, “Audio adversarial examples: Targeted at-
tacks on speech-to-text,” in 2018 IEEE Security and Privacy Workshops
(SPW), 2018, pp. 1–7.
[11] P. Cawley, M. Lowe, D. Alleyne, B. Pavlakovic, and P. Wilcox,
“Practical long range guided wave inspection-applications to pipes and
rail,” Mater. Eval, vol. 61, no. 1, pp. 66–74, 2003.
[12] S. Chen, K. Ren, S. Piao, C. Wang, Q. Wang, J. Weng, L. Su,
and A. Mohaisen, “You can hear but you cannot steal: Defending
against voice impersonation attacks on smartphones,” in 2017 IEEE 37th
International Conference on Distributed Computing Systems (ICDCS),
2017, pp. 183–195.
[13] M. Cisse, Y. Adi, N. Neverova, and J. Keshet, “Houdini: Fooling deep
structured prediction models,” arXiv preprint arXiv:1707.05373, 2017.
[14] H. Deng, W. Wang, and C. Peng, “Ceive: Combating caller id spoofing
on 4g mobile phones via callee-only inference and verification,” in
ACM International Conference on Mobile Computing and Networking,
MobiCom. ACM, 2018.
[15] W. Diao, X. Liu, Z. Zhou, and K. Zhang, “Your voice assistant is mine:
How to abuse speakers to steal information and control your phone,”
in Proceedings of the 4th ACM Workshop on Security and Privacy in
Smartphones & Mobile Devices, 2014, pp. 63–74.
[16] A. Dobrucki, “Nonlinear distortions in electroacoustic devices,”
Archives of Acoustics, vol. 36, no. 2, pp. 437–460, 2011.
17
[17] duo.com, “DUO,” https://duo.com/product/trusted-users/
two-factor-authentication/authentication-methods, 2019, accessed:
2019-09-07.
[18] V. Giurgiutiu, “Lamb wave generation with piezoelectric wafer active
sensors for structural health monitoring,” in Smart Structures and
Materials 2003: Smart Structures and Integrated Systems, vol. 5056,
2003, pp. 111–123.
[19] Y. Gong and C. Poellabauer, “Crafting adversarial examples for speech
paralinguistics applications,” arXiv preprint arXiv:1711.03280, 2017.
[20] ——, “An overview of vulnerabilities of voice controlled systems,”
arXiv preprint arXiv:1803.09156, 2018.
[21] Google, “Allow lock screen personal results on your android phone,”
https://support.google.com/assistant/answer/9134021, 2019, accessed:
2019-09-04.
[22] R. Gorgin, Z. Wu, D. Gao, and Y. Wang, “Damage size characterization
algorithm for active structural health monitoring using the a0 mode of
lamb waves,” Smart Materials and Structures, vol. 23, no. 3, p. 035015,
2014.
[23] K. F. Graff, Wave motion in elastic solids. Courier Corporation, 2012.
[24] T. Hayashi, W.-J. Song, and J. L. Rose, “Guided wave dispersion curves
for a bar with an arbitrary cross-section, a rod and rail example,”
Ultrasonics, vol. 41, no. 3, pp. 175–183, 2003.
[25] intellias, “The rise of voice payment technology in bankingh,” https://
www.intellias.com/the-rise-of-voice-payment-technology-in-banking/,
2019, accessed: 2019-09-07.
[26] Y. Jang, C. Song, S. P. Chung, T. Wang, and W. Lee, “Ally attacks:
Exploiting accessibility in operating systems,” in Proceedings of the
2014 ACM SIGSAC Conference on Computer and Communications
Security, 2014, pp. 103–115.
[27] P. Jayaram, H. Ranganatha, and H. Anupama, “Information hiding
using audio steganography–a survey,” The International Journal of
Multimedia & Its Applications (IJMA) Vol, vol. 3, pp. 86–96, 2011.
[28] C. Kasmi and J. L. Esteves, “Iemi threats for information security: Re-
mote command injection on modern smartphones,” IEEE Transactions
on Electromagnetic Compatibility, vol. 57, no. 6, pp. 1752–1755, 2015.
[29] X. Lei, G.-H. Tu, A. X. Liu, C.-Y. Li, and T. Xie, “The insecurity
of home digital voice assistants-amazon alexa as a case study,” arXiv
preprint arXiv:1712.03327, 2017.
[30] J. Liu, Y. Chen, M. Gruteser, and Y. Wang, “Vibsense: Sensing touches
on ubiquitous surfaces through vibration,” in 2017 14th Annual IEEE
International Conference on Sensing, Communication, and Networking
(SECON), 2017, pp. 1–9.
[31] J. Liu, C. Wang, Y. Chen, and N. Saxena, “Vibwrite: Towards finger-
input authentication on ubiquitous surfaces via physical vibration,” in
Proceedings of the 2017 ACM SIGSAC Conference on Computer and
Communications Security, 2017, pp. 73–87.
[32] K. Liu, S. Ma, Z. Wu, Y. Zheng, X. Qu, Y. Wang, and W. Wu, “A novel
probability-based diagnostic imaging with weight compensation for
damage localization using guided waves,” Structural Health Monitoring,
vol. 15, no. 2, pp. 162–173, 2016.
[33] Lyrebird, “Ultra-realistic voice cloning and text-to-speech,” https://
lyrebird.ai/, 2019, accessed: 2019-09-07.
[34] mcmaster, “Mcmaster-carr,” https://www.mcmaster.com/, 2020, ac-
cessed: 2020-01-10.
[35] M. Mitra and S. Gopalakrishnan, “Guided wave based structural health
monitoring: A review,” Smart Materials and Structures, vol. 25, no. 5,
p. 053001, 2016.
[36] G. Petracca, Y. Sun, T. Jaeger, and A. Atamli, “Audroid: Preventing
attacks on audio channels in mobile devices,” in Proceedings of the 31st
Annual Computer Security Applications Conference, 2015, pp. 181–190.
[37] T. L. Project, “Lineageos android distribution,” https://https://lineageos.
org/, 2020, accessed: 2020-01-03.
[38] N. L. Psesce, “Here’s why you’re getting so many
spam phone calls,” https://www.marketwatch.com/story/
heres-why-youre-getting-so-many-spam-phone-calls-2018-10-12,
2017, accessed: 2019-09-07.
[39] J. L. Rose, Ultrasonic waves in solid media. Cambridge university
press, 2004.
[40] N. Roy and R. R. Choudhury, “Ripple II: Faster communication through
physical vibration,” in 13th USENIX Symposium on Networked Systems
Design and Implementation (NSDI 16), 2016, pp. 671–684.
[41] N. Roy, H. Hassanieh, and R. Roy Choudhury, “Backdoor: Making
microphones hear inaudible sounds,” in Proceedings of the 15th An-
nual International Conference on Mobile Systems, Applications, and
Services, 2017, pp. 2–14.
[42] N. Roy, S. Shen, H. Hassanieh, and R. R. Choudhury, “Inaudible
voice commands: The long-range attack and defense,” in 15th USENIX
Symposium on Networked Systems Design and Implementation (NSDI
18), 2018, pp. 547–560.
[43] L. Sch
¨
onherr, K. Kohls, S. Zeiler, T. Holz, and D. Kolossa, “Adversarial
attacks against automatic speech recognition systems via psychoacoustic
hiding,” in Network and Distributed Systems Security (NDSS) Sympo-
sium, 2019.
[44] L. Song and P. Mittal, “Inaudible voice commands,” arXiv preprint
arXiv:1708.07238, 2017.
[45] STEMINC, “Piezo disc 92 khz r 22x0.25mm wire lead,” https://www.
steminc.com/PZT/en/piezo-disc-7-mhz-r-22x025mm-wire-lead, 2019,
accessed: 2019-09-07.
[46] Z. Su and L. Ye, Identification of damage using Lamb waves: from
fundamentals to applications. Springer Science & Business Media,
2009, vol. 48.
[47] T. Trippel, O. Weisse, W. Xu, P. Honeyman, and K. Fu, “Walnut:
Waging doubt on the integrity of mems accelerometers with acoustic
injection attacks,” in 2017 IEEE European Symposium on Security and
Privacy (EuroS&P), 2017, pp. 3–18.
[48] T. Vaidya, Y. Zhang, M. Sherr, and C. Shields, “Cocaine noodles:
exploiting the gap between human and machine speech recognition,” in
9th USENIX Workshop on Offensive Technologies (WOOT 15), 2015.
[49] wiki, “Absolute threshold of hearing,” https://en.wikipedia.org/wiki/
Absolute\
threshold\ of\ hearing, 2020, accessed: 2020-01-17.
[50] P. Wilcox, M. Lowe, and P. Cawley, “Mode and transducer selection
for long range lamb wave inspection,” Journal of intelligent material
systems and structures, vol. 12, no. 8, pp. 553–565, 2001.
[51] X. Yuan, Y. Chen, Y. Zhao, Y. Long, X. Liu, K. Chen, S. Zhang,
H. Huang, X. Wang, and C. A. Gunter, “Commandersong: A systematic
approach for practical adversarial voice recognition,” in 27th USENIX
Security Symposium (USENIX Security 18), 2018, pp. 49–64.
[52] G. Zhang, C. Yan, X. Ji, T. Zhang, T. Zhang, and W. Xu, “Dolphi-
nattack: Inaudible voice commands,” in Proceedings of the 2017 ACM
SIGSAC Conference on Computer and Communications Security, 2017,
pp. 103–117.
[53] N. Zhang, X. Mi, X. Feng, X. Wang, Y. Tian, and F. Qian, “Dangerous
skills: Understanding and mitigating security risks of voice-controlled
third-party functions on virtual personal assistant systems,” in IEEE
Symposium on Security and Privacy (SP), 2019.
APPENDIX
A. Material Properties
The material properties of five different materials, including
Ceramic, Aluminum, Glass, MDF, HDPE, are listed in Table X.
TABLE X: Material properties and acoustic impedances of
different materials.
Materials
Young’s module
(Gpa)
Poisson ratio
Density
(kg/m
3
)
Acoustic impedance
(MPa· s/m)
Ceramic 350 0.22 3,800 23.3
Aluminum 70 0.33 2,740 8.5
Glass 50 0.22 2,580 7.27
MDF 4 0.25 750 1.1
HDPE 0.6 0.46 930 0.437
18
