2023 IEEE the 7th Conference on Energy Internet and Energy System Integration

CamDetector: Detecting and Positioning Hidden

Cameras with Raspberry Pi 4 via EM Radiation
Han Chen Ruochen Zhou
College of Electrical Engineering College of Electrical Engineering
Zhejiang University Zhejiang University
Hangzhou, Zhejiang Hangzhou, Zhejiang
[email protected] [email protected]
2023 IEEE 7th Conference on Energy Internet and Energy System Integration (EI2) | 979-8-3503-4509-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/EI259745.2023.10512830

Chen Yan Xiaoyu Ji Wenyuan Xu

College of Electrical Engineering College of Electrical Engineering College of Electrical Engineering
Zhejiang University Zhejiang University ZheJiang University
Hangzhou, Zhejiang Hangzhou, Zhejiang Hangzhou, Zhejiang
[email protected] [email protected] [email protected]

Abstract—The illicit use of spy cameras poses a significant

threat to scenarios where privacy is paramount, such as in Sensor SoC
hotel rooms and home rentals. Owing to advancements in
miniaturization and camouflage techniques, these covert cameras
have become increasingly difficult for victims to detect in their
surroundings. In this paper, we introduce a proof-of-concept EM Radiation
system called CamDetector, designed to identify spy cameras
based on their electromagnetic (EM) radiations. Our initial step Preliminary Feature Camera
involves characterizing the EM radiation emitted by cameras. Processing Judgement Verification
We then uncover the shared pattern stemming from the signal CamDetector
coupling on the CSI cable, which can be used to detect hidden
cameras. We use USRP B200 mini and Raspberry Pi 4 to
build the prototype and proceed to assess the performance on 6 Fig. 1. CamDetectoraims at detecting hidden spy cameras via the EM
radiation, which is generated when data is transmitted between the sensor and
commercially available small cameras. The experimental results
SoC. By preliminary processing and feature judgement, CamDetectorobtains
demonstrate the effectiveness of CamDetectorwith an average unique EM patterns of cameras and completes detection.
accuracy of 90.08% and a recall of 88.21% at a distance of 0.4m.
Moreover, our evaluation based on user involvement confirmed
that CamDetector can achieve 85% accuracy in an 18 m2 room,
with an average time of 16.9 minutes. an effective method of detecting spy cameras, to ensure their
Index Terms—Electromagnetic Radiation, Hidden Camera, privacy.
Device Detection Detecting a spy camera is a non-trivial task because video
recording is fundamentally a passive process. Unlike active
I. I NTRODUCTION sensors, camera sensors are not designed to emit signals into
Camera spying poses a significant threat to personal privacy. the environment; instead, they passively measure light inten-
There have been numerous cases where victims discovered sity using pixel arrays and convert this information into elec-
hidden pinhole cameras in hotel rooms. With the prolifer- trical signals for subsequent image processing. Additionally,
ation of Complementary Metal-Oxide-Semiconductor Tran- many cameras have offline functions that allow them to save
sistor (CMOS) sensors, cameras have become increasingly videos locally on SD cards, meaning they do not necessarily
miniature, making it easier for them to be concealed within initiate wireless connections. Consequently, existing traffic-
everyday items, such as lamps, smoke detectors, routers, and based solutions [3], [4] are insufficient for detecting offline
power outlets. This, in turn, makes it more challenging for spy cameras. Other methods of detecting spy cameras, e.g.,
victims to detect them. by lens reflection [5], thermal radiation [6], power traces [7],
The rise in popularity of home rentals, such as Airbnb and or illumination [8], [9], suffer a higher false positive rate due
VRBO, has also exacerbated this issue. These shared-space to the presence of interfering devices.
scenarios create opportunities for adversaries to secretly place Despite these challenges, this paper tries to detect hidden
spy cameras in advance, intensifying privacy concerns. In spy cameras by finding the unique characteristics of their EM
response, 15 states in the United States enact specific laws to radiations. We first find there are lower-frequency EM radia-
restrict the use of cameras in private places, e.g., bedrooms and tion peaks that appear around the higher-frequency peaks with
bathrooms [1], [2]. Nevertheless, people are always seeking an equal interval in the spectrum, a result we call adjacent-

979-8-3503-4509-4/23/$31.00 ©2023 IEEE 3597

Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.
signal coupling from the circuits. After a detailed investigation, MIPI Alliance, originally introduced in 2005, is the world’s
we find the lower-frequency EM radiation derives from the most widely implemented embedded camera and imaging
pixel readout process which is performed line by line (defined interface. It has achieved widespread adoption for its ease of
as Line Readout in the following), typically tens of kilohertz use and ability to support a broad range of high-performance
and the higher-frequency EM radiation derives from the clock applications, including 1080p, 4K, 8K and beyond video, and
signal used by CMOS sensors. This finding makes it possible high-resolution photography.
to detect hidden cameras using RF signal signatures. The CSI can be divided into two parts: camera control
Based on the above findings, we propose our detection interface and data transmission interface. The former contains
system, CamDetector, a proof-of-concept system that can SCL and SDA using I2C protocol, the latter contains a
detect spy cameras by their EM radiations. The detection differential clock lane and up to four differential data lanes.
system consists of two stages. Firstly sweeps the frequency When transmitting pixel data of one line, it consists of two
from 100 MHz to 500 MHz to get the full-band spectrum. stages which are low power state (LPS) and High-Speed
Secondly determines the feature of frequency peaks both in Transmission (HSPTX). Note that the pixels from camera
the frequency domain and time domain. sensors are transmitted line by line, the period T of line
We deploy our prototype with USRP B200 mini and Rasp- readout signal is determined by the frame rate and the number
berry Pi 4. The time overhead of one-round detection is of lines, typically tens of kilohertz.
around 11.3 s. Our evaluation in hotel rooms on 6 commer-
B. EM Radiation from Electrical Devices
cial small cameras under the interference of other electronic
devices shows that CamDetectorcan effectively detect all of Maxwell’s equations and the Lorentz force law provide
the cameras when they are hidden with an average success the foundation for understanding how charges, currents, and
rate of 90.08% and a recall rate of 88.21% at a distance of alterations in electric and magnetic fields generate these fields.
0.4 m. Moreover, our evaluation also shows that CamDetec- Additionally, they elucidate the propagation of fluctuating
torachieves efficient localizability with an accuracy of about electric and magnetic fields through space, a phenomenon
85% within 16.9 minutes. referred to as EM radiation. Electrical devices, encompassing
In summary, this paper makes the following contributions. digital, analog and power circuits, inherently involve time-
varying current flows, resulting in the emission of diverse EM
• We find the common EM radiation patterns emitting from
radiations.
CSI cable which can be used for detecting spy cameras.
Considering a clock signal with its frequency fc , due to the
• We design CamDetector, a proof-of-concept system that
non-ideal and non-linear properties of circuits, the generated
can detect spy cameras based on weak EM radiations.
EM radiation has not only a fundamental frequency component
Our evaluation of 6 commercial cameras in the lab and
but also its harmonic components at the n · fc , where n is
real-world cases demonstrates its effectiveness.
an integer. In addition, existing work [10] has demonstrated
• We further explore the ability of CamDetectorto localize
that multiple signal sources will be radiated in the form of
the hidden cameras and verify its efficiency through real-
“carrier coupling” if they are spatially adjacent signal circuits.
world user-involvement experiments.
Therefore, if there are two strong carriers with frequencies fc1
The rest of this paper is organized as follows. Section II and fc2 in a circuit, the coupled signal radiated frequency fr
provides a brief background on EM radiation from electrical can be expressed as,
devices and the workflow of cameras. Section III reports the
procedure in that we explore the EM radiation of cameras. fr = n · fc1 ± m · fc2 n, m ∈ N (1)
Section IV and Section V design and evaluate CamDetector.
Finally, we conclude this paper in Section VI. This paper aims to investigate the shared EMR patterns of
cameras and assess their potential for camera detection.
II. BACKGROUND
III. EMR PATTERN OF C AMERAS
A. Workflow of Cameras To find the unique spectrum pattern of cameras, we use
A camera system’s workflow typically includes three key the Raspberry Pi camera v2 for experiments, which can easily
procedures: capturing, transmission, and processing. As shown adjust the parameters.
in fig. 1, an image sensor (mostly CMOS sensors) captures and While the cameras are working offline, we adopt a near-field
converts ambient light to digital signals which are transmitted probe to capture the RF signals in close range to the cameras.
through a standard camera interface to an ISP, short for image The probe is connected to a low-noise amplifier (LNA) to
signal process, that further processes the raw image or video. amplify weaker signals, and then software-defined radio (SDR)
Several serial and parallel buses have been proposed for is used to down-convert and digitize the amplified signals. A
public and proprietary use to connect an image sensor and an laptop in the end is used to analyze the spectrum of received
image processor, such as DVP. MIPI, short for Mobile Industry signals and perform signal processing methods. To acquire
Processor Interface, is a set of specifications developed by the a fine-grained spectrum, we sweep a wide-band signal that
MIPI Alliance, a global consortium of technology companies. ranges from 50 MHz to 1 GHz with a step of 20 MHz, and
CSI-2, short for Camera Serial Interface 2, developed by the sampling rate of SDR is set to 20 MHz.

3598
Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.
 (a) Frame rate (15 Hz) (b) Line readout waveform (52.1 kHz)

Fig. 3. Signal waveform of one of data lines on the CSI pinouts

(a) full-band sepctrum of Raspberry Pi V2 camera

Fig. 4. Magnitude waveform of baseband frequency

(b) Grained spectrum around 225MHz of Raspberry Pi V2 camera

Fig. 2. The spectrum of the EM Radiation emanated from Raspberry Pi V2

camera. The figure above represents the full-band spectrum and harmonics with an interval between every two adjacent frames and the
with red triangle label. The figure below represents the fine-grained spectrum. current frame rate is 15 fps. After scaling up the x-axis, we
can observe the process of line-by-line transmission within
A. Spectrum Patterns a frame. The waveform can be approximately viewed as a
square wave with a frequency of 52.1 kHz, which is roughly
Fig.2 shows the power spectrum of Raspberry Pi v2 camera.
consistent with 53.1 kHz, the frequency of baseband signal.
From the result, we derive two observations about the EM
Moreover, the frequency of baseband’s magnitude is same as
radiation of the camera.
the camera’s framerate, shown in the fig. 4. It makes sense
• The peaks on the spectrum can be grouped into harmonics here, there is no data transmitting between frames thus leading
shown in fig. 2(a), most likely generated by the system to lower magnitude. At this point, we can conclude that the
clock. We regard these signals as carrier. camera’s special spectral pattern is due to the signal coupling
• There are lower frequency signals coupling to one of of clock line and data line on CSI cable.
harmonics, shown in fig. 2(b). We regard these signals The display works similar to the camera, but there is almost
as baseband. no EM radiation due to the aluminum layer in HDMI and VGA
This phenomenon is consistent with eq. (1) we proposed cables. Therefore, this pattern is unique to cameras and can
before. The fundamental frequency of the harmonics we got be used to detect hidden cameras.
is 24 MHz and the lower frequency interval is 53.1 KHz. We
also tested other cameras and got similar results. IV. SYSTEM DESIGN
Through these results, we can preliminarily conclude that As outlined in fig. 1, CamDetectortakes RF signal as input
EMRs of cameras do have common spectral features. Next, and reports the presence of cameras in the surrounding envi-
we will prove that this feature is unique to cameras. ronment. The details of two main stages are described in the
following subsections.
B. Source of EM Radiation
By referring to the datasheet, we find that 24MHz is the A. Full-band Sweeping and Preprocessing
system clock frequency of the v2 camera. The clock signal We use SDR to capture the RF signal and focus on 100 MHz
serves as the synchronization signal for the entire system, to 500 MHz, the bandwidth where the camera’s RF signals
whose high intensity makes it easy to serve as the carrier mainly concentrates on. We set sample rate to 10MHz so that
signal. The occurrence of signal coupling requires that the devices like LimeSDR and HackRF that are cheep can meet
carrier signal and the baseband signal be very close to each this requirement and Raspberry Pi 4 can meet the computing
other [11], [12]. This naturally leads to the conjecture that power needs. Through our experiments, the lowest sample rate
this phenomenon occurs on the CSI transmission line where is 5MHz.
the clock line and data line are close to each other and the There are two kinds of noise in the captured signal: white
cable can work as an antenna which benefits the radiation. noise in receiving circuit and EM noise from other electronic
We verify this by measuring the waveform of data line with devices. The former can be removed by averaging a set of
an oscilloscope and the results are shown in fig. 3. It can samples, and the latter can be removed by sampling RF
be clearly seen that the video is transmitted frame by frame signal where far away from target cameras due to its stable

3599
Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.
 (a) Before folding (b) After folding

Fig. 5. The line frequency coupled at 381.825MHz before and after folding
at 30 cm.

characteristics. The spectral subtraction proposed in [13]

is used to enhance the intelligence of speech corrupted by
acoustic noise, it also works in clean the frequency domain
by set α to 1.2 and β to 0 in eq. (2).
(
S(f ) − αD(f ) S(f ) > (α + β)D(f )
S0 (f ) = (2) Fig. 6. The appearance of the 6 chosen cameras, #C1 - #C2 are battery-
βD(f ) else
powered cameras, and #C3 - #C6 are USB-powered cameras.
In fact, we average 10 samples each contains 16384 points
with 10MHz bandwidth to reduce white noise and build the TABLE I
full-band spectrum without cameras by averaging 40 samples. OVERALL PERFORMANCE OF 6 SPY CAMERAS .

This stage will cost time around 2s, we can assume the peak
Max Distance.
value is emitted from hidden cameras after subtraction. Label Power
(cm)
B. Feature Judgement C1 Battery 50
C2 Battery 43
After we get the full-band spectrum, we will focus on if the C3 USB 88
spectrum has the feature mentioned in section III. C4 USB 57
1) Frequency Domain: First, we should group the harmon- C5 USB 132
C6 USB 165
ics of system clock used by CMOS sensor for further analysis.
We use the folding method to better group the harmonics
and improve the SNR by presetting the period size and
amining the max detection distance of cameras in laboratory
segmentation, and comparing the peaks after superimposing.
environment and examing the SNR and TPR at various dis-
We learn that the commercial cameras mostly have resolu-
tances. Following that, we evaluate the detection accuracy of
tions of 1920*1080, 1080*720 and frame rate ranging from
CamDetectorin a real-world scene with interference from other
10fps to 30fps. So the line frequency fl is in the range of
devices. Finally, we invite volunteers to use CamDetectorto
10KHz to 60KHz by calculating. This value can be used to
detect hidden cameras in a room and evaluate their detection
jointly determine the feature of the frequency domain. After
time and accuracy.
folding each component of harmonics, the line frequency
can also be grouped as shown in Fig 5(a)-(b). The SNR of A. Experimental Setup
baseband frequency improves from 7.2dB to 21.6dB which
benifits the detection distance. Targeted cameras. Our criteria for a targeted camera are
2) Time Domain: We can further improve the detection small-size and ease of disguise, i.e., more likely to be used
accuracy by observing the feature of time domain mentioned as a spy camera by an attacker. fig. 6 shows the 6 cameras
in section III-B as the magnitude of baseband’s frequency is we selected on top shopping platforms, consisting of 2 com-
consistent with frame rate of the hidden camera. To minimize mercially available battery-powered compact cameras and 4
the time of one detection, we choose 1.6 s to record the USB-powered cameras. We present the label, power supply
magnitude. In our experiments, the average time of detection type and corresponding results in table I.
is 11.3 s. Equipment. As shown in fig. 6, A to F stand for DC power
supply, a magnetic filed probe NFP-3, a low noise amplifier
V. E VALUATION LNA-650, the USRP B200 mini, a Raspberry Pi 4 and a 3.5
In this section, we describe our experimental setup and inch display. These devices are small, inexpensive and easy to
evaluate the performance of CamDetector. We start by ex- carry.

3600
Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.
Fig. 7. The relationship between the average TPR and the detection distance.

Metrics. We adopt the following metrics throughout the

Fig. 8. TPR and TNR of 6 cameras in real world.
evaluation. As a premise, we consider recording spy cameras
as the “positives” for our system, and any device else should
A
be “negatives”. 1
L
• True Positive Rate (TPR): characterizes the probability C

that the system correctly detects the positive samples.

• True Negative Rate (TNR): characterizes the probability 1 2
B
that the system correctly ignores the negative samples. D
• Accuracy & Recall:
(a) Floor Plan of the 18m2 room (b) Localization via magnitude
change
TP + TN TP
Accuracy = P , Recall = (3) Fig. 9. The experiments setup under real-world scenes. (a) We set CamDe-
samples TP + FN tectorat the orange spot, and the 6 cameras in fig. 6 are hidden at the red spot
B. Overall Performance while testing distance and green spots while localizing. (b) The magnitude
change with user’s move.
In laboratory environment, We conduct 20 times detection
for each camera at each distance that ranges from 1 cm to
50 cm with a step length of 1 cm. table I summarizes the and 50 cm. We also provide the average TPR of 6 cameras
results for each camera. and TNR. The results indicate that CamDetectorkeeps a good
We can find that the average maximum detection distance performance in real-world cases, with an accuracy of 90.08%
of #C3 - #C6 (USB-powered) is much larger than that of (78.50%) and recall of 88.21% (74.20%) at 40 cm (50 cm).
#C1 - #C2 (battery-powered). We guess this is because the
cameras powered by USB do not need to meet the low D. Camera Location
power consumption requirements, resulting a longer detection
distance. The magnitude of baseband signal increases as the CamDe-
To illustrate the detection performance under various dis- tectorgets closer to the camera, and vice versa. Therefore,
tances, fig. 7 shows the average TPR of 6 cameras at various we can locate the hidden cameras by the gradient changes
distances. The maximum detection distance of all cameras of magnitude. After the detector’s circle in the room as
achieves more than 40 cm while the largest one can even reach shown in fig. 9(b), we get the place with the largest detection
165 cm. CamDetectorachieves an average TPR of 87.73% at a signal magnitude, the brown spot. It can be seen that it’s
distance of 40 cm and 73.64% at 50 cm. As a comparison, the close to the red spot where the spy camera is deployed.
existing state-of-art detection system, CamRadar [14], claims Actually, in the event that the hidden camera has been detected,
a TPR of 70.68% at the farthest distance of 40cm. CamDetectorwill also tell the user whether he is closer or
further away from it by his movement. We invite 20 volunteers
C. Real-world Cases in our college, half male and half female, to test the ability of
To investigate the practical performance of CamDetector, CamDetectorto locate the camera.
we conduct further experiments in a real-world case, an 18m2 We start by guiding them on how to use the system, and
hotel room. The floor plan of the room is shown in fig. 9(a), telling them that there is a hidden spy camera with an unknown
and we hide each camera in fig. 6 at the red spot ‘1’, i.e., location in the room in fig. 9(a). Similarly, we implant the spy
underneath the TV, and CamDetectoris implemented at the cameras around the green spots and A to D stand for opposite
orange spot 2, keeping the distance between the probe and the bed, on top of the cabinet, in the bathroom and next to the
TV about 40-50 cm. For each of the 6 cameras, we test 20 TV respectively. The goal of participants is to scan around the
times, respectively. In addition, we also conduct 100 rounds room, follow the instructions by CamDetector, and try to find
of tests while spy cameras are off to evaluate the TNR of the out the hidden camera. Finally, the results are recorded in the
system. table II, and show that 17 out of 20 participants successfully
In fig. 8, we present the TPR and TNR of 2 battery-powered find the spy camera without false positives, and it takes an
cameras, 4 USB-powered cameras under distances of 40 cm average of 16.9 minutes.

3601
Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.
 TABLE II [14] Z. Liu, F. Lin, C. Wang, Y. Shen, Z. Ba, L. Lu, W. Xu, and K. Ren,
L OCATION RESULTS OF C AM D ETECTOR “Camradar: Hidden camera detection leveraging amplitude-modulated
sensor images embedded in electromagnetic emanations,” Proceedings of
Position Number Time Accuracy the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies,
A 5 14.6min 100% vol. 6, no. 4, pp. 1–25, 2023.
B 5 15.4min 100%
C 5 20.3min 60%
D 5 17.2min 80%

VI. CONCLUSION

This paper presents a state-of-art method for detecting spy

cameras in shared spaces. We identify both common and
distinct EM patterns arising from the coupling of the clock
signal with the line readout signal, forming the basis for
spy cameras detection. We develop CamDetector, a proof-
of-concept system designed for this purpose. Our evaluation
demonstrates that CamDetectorattains 90.08% accuracy at a
distance of 40 cm and can also localize hidden cameras with
an accuracy of approximately 85%, all within a reasonable
time frame.

R EFERENCES
[1] Rebecca Edwards, “Security camera laws, rights, and rules,” 2022,
https://www.safewise.com/security-camera-laws/.
[2] Aliza Vigderman, Gabe Turner, “Are security cameras legal?” 2022,
https://www.security.org/security-cameras/legality/.
[3] Y. Cheng, X. Ji, T. Lu, and W. Xu, “Dewicam: Detecting hidden
wireless cameras via smartphones,” in Proceedings of the 2018 on Asia
Conference on Computer and Communications Security, 2018, pp. 1–13.
[4] Y. He, Q. He, S. Fang, and Y. Liu, “Motioncompass: pinpointing wireless
camera via motion-activated traffic,” in Proceedings of the 19th Annual
International Conference on Mobile Systems, Applications, and Services,
2021, pp. 215–227.
[5] S. Sami, S. R. X. Tan, B. Sun, and J. Han, “Lapd: Hidden spy camera
detection using smartphone time-of-flight sensors,” in Proceedings of the
19th ACM Conference on Embedded Networked Sensor Systems, 2021,
pp. 288–301.
[6] A. Zuniga, N. H. Motlagh, M. A. Hoque, S. Tarkoma, H. Flores, and
P. Nurmi, “See no evil: Discovering covert surveillance devices using
thermal imaging,” IEEE Pervasive Computing, vol. 21, no. 4, pp. 33–42,
2022.
[7] M. Jin, R. Jia, Z. Kang, I. C. Konstantakopoulos, and C. J. Spanos, “Pres-
encesense: Zero-training algorithm for individual presence detection
based on power monitoring,” in Proceedings of the 1st ACM conference
on embedded systems for energy-efficient buildings, 2014, pp. 1–10.
[8] H. Aniktar, D. Baran, E. Karav, E. Akkaya, Y. S. Birecik, and M. Sezgin,
“Getting the bugs out: A portable harmonic radar system for electronic
countersurveillance applications,” IEEE Microwave Magazine, vol. 16,
no. 10, pp. 40–52, 2015.
[9] J. Raoult, A. Martorell, L. Chusseau, and C. Carel, “Intermodulation
radar for rf receiver detections,” in Proceedings of the 2018 15th
European Radar Conference (EuRAD). IEEE, 2018, pp. 273–276.
[10] P. Lui, “Passive intermodulation interference in communication sys-
tems,” Electronics & Communication Engineering Journal, vol. 2, no. 3,
pp. 109–118, 1990.
[11] G. Camurati, S. Poeplau, M. Muench, T. Hayes, and A. Francillon,
“Screaming channels: When electromagnetic side channels meet radio
transceivers,” in Proceedings of the 2018 ACM SIGSAC Conference on
Computer and Communications Security, 2018, pp. 163–177.
[12] J. Choi, H.-Y. Yang, and D.-H. Cho, “Tempest comeback: A realistic
audio eavesdropping threat on mixed-signal socs,” in Proceedings of
the 2020 ACM SIGSAC Conference on Computer and Communications
Security, 2020, pp. 1085–1101.
[13] M. Berouti, R. Schwartz, and J. Makhoul, “Enhancement of speech cor-
rupted by acoustic noise,” in ICASSP’79. IEEE International Conference
on Acoustics, Speech, and Signal Processing, vol. 4. IEEE, 1979, pp.
208–211.

3602
Authorized licensed use limited to: Malnad College of Engineering. Downloaded on May 22,2025 at 05:48:40 UTC from IEEE Xplore. Restrictions apply.