Under Pressure: Pushing Down on Me – Touch Sensitive Door Handle to
Identify Users at Room Entry
Christian Tietz, Eric Klieme, Rachel Brabender, Teresa Lasarow, Lukas Rambold
and Christoph Meinel
Hasso Plattner Institute, University Potsdam, Potsdam, Germany
Keywords:
Behavior Biometrics, Identification, Door Handle, Touch Interactions.
Abstract:
Each day we open a door using physical keys or tokens like RFID or smart cards. While we all are used
to these methods they have problems of security and usability. These tokens and keys can easily be stolen
or taken by other persons which results in a security problem. The problem in usability is that users need a
significant amount of time to take out their tokens to unlock and open the door. In this paper, we propose a
new approach for door handle access control. We developed a prototype by attaching pressure sensors to the
door handle that measure resistive and capacitive touch interactions with the door handle. We demonstrate
the feasibility of identification with a door handle with a visual and classification analysis. The classification
algorithms used are K-NN, SVM, Random Forests, AdaBoost and MLP achieving a maximum accuracy of
88% using the Random Forests.
1 INTRODUCTION
Currently, authentication at door and gate access
points is mostly achieved by using mechanical keys
or radio-frequency tags (RFID). Each verification key
permits users to access associated areas. The prob-
lem with such physical tokens is that they can eas-
ily be cloned or stolen. For example, one can find a
wide variety of instructions to duplicate mechanical
keys without a locksmith (Marsh et al., 2014). Fur-
thermore, there exist smartphone apps that let people
easily scan their physical keys with the camera, store
it in the cloud or share them with family & friends
and easily order duplicates that are sent to you by mail
(Wendt, 2015). Also, RFID tags are not secure as re-
search showed that it was possible to open millions
of doors without authorization in a hotel (Pinkert and
Tanriverdi, 2018) or clone a key of a Tesla Model S
(Greenberg, 2018).
To make access control more secure, biomet-
rics can be used. For example, fingerprint (Odiete
et al., 2017), palmprint (A. Kumar and Jain, 2003),
hand contour (Schmidt et al., 2010), voice (Wahyudi
and Syazilawati, 2007), face recognition (Alam and
Yeasin, 2019)(Varasundar and Balu, 2015) and com-
binations of them (Brunelli and Falavigna, 1995)(Bi-
gun et al., 2005) are ways to unlock a door. These
methods are also having some problems. Fingerprints
can be photographed and forged (ChaosComputer-
Club, 2013) or iris recognition can be bypassed with
a simple photo (ChaosComputerClub, 2017).
Both, the possession- and biometric-based meth-
ods have a usability problem. All these methods re-
quire an additional authentication effort besides open-
ing the door, e.g. interacting with the lock or the bio-
metric scanner terminal. Recent research showed that
users like the idea of just using the door handle with-
out any additional interaction (Mecke et al., 2018).
They compared physical keys, a gait-based system,
and a vein scanner integrated into the door handle as
methods in a Wizard-of-Oz study and analyzed the
perception of users. The results showed that users
like seamless interaction with the vein scanner and the
door handle because it is faster than a key, more com-
fortable, easy to use and secure.
In this work, we propose a proof of concept for
physical access control based on the user’s behavior
while using the door handle. A normal door handle is
enhanced with resistive and capacitive pressure sen-
sors instead of a real vein scanner. Using touch sen-
sors is new in the field of door handle access control.
We give a summary of existing research in this
area in Section 2. Then, we present our door han-
dle prototype and the data collection approach in Sec-
tion 3 and 4. Afterwards, we go over the data ex-
traction and pre-processing (Section 5), followed by a
first data exploration in Section 6. We finish the paper
with an evaluation of the identification results and the
conclusion (Sections 7 and 8).
Tietz, C., Klieme, E., Brabender, R., Lasarow, T., Rambold, L. and Meinel, C.
Under Pressure: Pushing Down on Me – Touch Sensitive Door Handle to Identify Users at Room Entry.
DOI: 10.5220/0009818805650571
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - SECRYPT, pages 565-571
ISBN: 978-989-758-446-6
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
565
2 RELATED WORK
In the field of access control using door handles, there
is already some prior work done. We divided them
into two parts: image-based and sensor-based evalua-
tion.
2.1 Door Handle Authentication using
Images
Aoyama et al. (Aoyama et al., 2013) proposed a sys-
tem for analyzing the user’s finger knuckles on the
door handle. The door is set up with a camera and
an infra-red-light source that records an image of the
four-finger knuckles while interacting with the door
handle. From the hand image, they detect and ex-
tract the knuckles as ROI (region of interest). They
used 900 images from 90 subjects (10 images per per-
son, 5 left, and 5 right hands). The algorithms used
are knuckling recognition methods: BLPOC, pCode
and LGIC and their proposed one. The best result is
achieved by middle and ring finger combination with
an EER (Equal-Error-Rate) of 1.54%.
A similar approach was presented by Kusanagi at
al. (Kusanagi et al., 2017). They used a camera to get
the images of the finger knuckles from above and not
from the front. They also extracted the finger knuck-
les ROI from the image. Their database was created
from 28 participants who also used both, left and right
hand 5 times each. This for two sessions. In a total of
560 images. They also came up with a new proposal
and compared them to the existing BLPOC, Comp-
Code and LGIC algorithms. They evaluated each fin-
ger individually and in combination. The best result
is achieved with a combination of all four knuckles
resulting in an EER of 2.36%.
2.2 Door Handle Authentication using
Sensors
Garcia et al. (Garcia et al., 2016) investigated hand
dynamics and the door handle movement when open-
ing a door. They used two smartphones that are at-
tached to the hand and the door handle to collect
data from 20 participants. Each participant opened
the door ten times. They extracted 170 features (85
from hand and 85 from door handle) with statistical
(e.g. mean, median, root mean square level) and phys-
ical (e.g. movement intensity, dominant frequency
energy) features. The authentication algorithm was
SVM with 92% accuracy to identify users using a 50-
50 train test split.
Ishida et al. (Ishida et al., 2017) looked into the
door opening and closing of a refrigerator. They at-
Figure 1: The final protoype from our door handle with the
lengthwise attached sensor stripes. The positions of the sen-
sor stripes are Top, Bottom, Back and Front.
tached pressure, accelerometer and gyroscope sensors
to the door handle. Their features are acceleration,
angular velocity and pressure values of gripping the
handle. On seven participants, an accuracy of 91.9%
is reached.
2.3 Summary
There are already approaches in using the door han-
dle behavior to identify and authenticate users that
work quite well by analyzing finger-knuckles or us-
ing smartphone sensors. The finger-knuckle approach
requires a camera that needs to be integrated into or
above the door. This makes it complicated for a more
realistic, real-world user study. When using sensors, a
complete smartphone was attached to the door handle
which has an impact on how users are gripping and
using the door handle.
Thus, in this work, we build a new prototype that
attaches pressure sensors to the door handle and can
easily be installed to real doors to analyze the user’s
behavior by using their pressure on the door handle.
3 TOUCH-SENSITIVE DOOR
HANDLE PROTOTYPE
Our prototype uses four silicon-based sensor stripes
of Tacterion
1
. These stripes measure touch (capaci-
tive) and applied force (resistive) data. The material
and their size of 90x9 mm makes them suitable for un-
even surfaces like a cylindrical door handle. In total,
the sensors produce eight values per reading.
All four sensors are attached to a door handle on
top (To), on the front (Fr), the back (Ba) and the bot-
tom (Bo) of the door handle and are fixed by regular
masking tape. It has a silver color to keep the genuine
look to not disturb participants. The resulting proto-
type is shown in Figure 1.
1
https://www.tacterion.com/development-kit
SECRYPT 2020 - 17th International Conference on Security and Cryptography
566
4 DATA COLLECTION
To our knowledge, no similar work has been done and
no available data set can be used. Therefore, we de-
scribe our user study and some preliminary consider-
ations in this section.
4.1 Preliminary Considerations
There are a lot of different factors that might influ-
ence our behavior of opening a door. These factors
can be approaching the door from different directions
like from left, from right or frontal. The position of
the door handle (left or right) the used hand or open-
ing the door by pushing or pulling might be important,
too. The door opening can also change over multiple
days or if people are emotional, angry or in a hurry.
Users might talk to other people, having something
in their hands or using different types of door han-
dles. All these can influence the behavior and bring
randomness to it.
For this proof of concept work, we decided to use
the most common factors. Participants will approach
the door from frontal and opening it by pushing us-
ing their preferred hand. The recording is done in
a supervised manner without distractions and special
emotions.
4.2 User Study
We conducted a supervised user study to record their
door handle behavior using the most common param-
eters discussed in Section 4.1. Each participant got
an explanation of the study’s purpose and procedure
which has to be agreed and signed in a consent form.
Afterward, they follow the procedure shown in Figure
2.
The participant starts outside the room and takes
a piece of a puzzle that was prepared in advance (1).
The puzzle serves as distraction task. Then, they walk
inside (2) and put the piece of the puzzle at the cor-
rect position inside(3). Finally, they leave the room
by pushing the door open(4). At this time, we record
their door handle behavior. This completes one round
which is repeated 30 times. The participants answer
a questionnaire after finishing all rounds. We asked
for information about gender, their preferred hand that
could correlate to the recorded data and their thoughts
and views about this authentication concept in the
questionnaire. 25 people took part in the study.
1
2
3
4
Figure 2: The running direction of the user study. The data
was collected when the users leave the room and push the
door.
5 DATA EXTRACTION AND
PRE-PROCESSING
Our prototype generates data with a frequency of
30Hz which results in 8x30 data points per second.
The data were recorded and stored per user with
no separation of stand by phases and door opening
phases.
To detect the door handle usages in the stream of
data, we use the resistive sensors values. They have a
clear zero line when there is no interaction in compar-
ison to the capacitive sensors that are always showing
low non-zero values. The values of the resistive sen-
sors are numbers in the interval [0, 4095]. For a given
raw signal, we assign to each timestamp a ”0” when
all resistive sensors’ values are under 500 and ”1”
if at least one is over 500. Afterward, sequences of
consecutive ”1”s are combined into intervals (see Al-
gorithm 1) with the index of the first and last ”1” as
interval borders.
Algorithm 1: Building intervals of consecutive 1s.
function BUILDINTERVALS(list) list of 0, 1
intervals ← []
start ← 0
for all i in list do
if i + 1 = |list| or list[i] 6= list[i + 1] then
if list[i] = 1 then
intervals ← intervals +[(start, i)]
end if
start ← i + 1
end if
end for
return intervals
end function
Under Pressure: Pushing Down on Me – Touch Sensitive Door Handle to Identify Users at Room Entry
567
If two neighboring intervals are very close to each
other (< 100 timestamps, roughly 3 seconds) then the
two intervals are fused into one as described in Algo-
rithm 2.
Algorithm 2: Fusing nearby intervals.
function FUSEINTERVALS(intervals)
fusedIntervals ← []
lastInterval ← intervals[0] tuple
i ← 1
while i < |intervals| do
(start, end) ← intervals[i]
if |start − lastInterval[1]| < 100 then
lastInterval = (lastInterval[0], end)
else
fusedIntervals.append(lastInterval)
lastInterval ← (start, end)
end if
if i = |intervals| − 1 then
fusedIntervals.append(lastInterval)
end if
i ← i + 1
end while
return fusedIntervals
end function
Finally, each interval is extended by two seconds by
subtracting 30 time-units from the beginning index
and adding 30 time-units to the end. These final inter-
vals are then used to extract the door opening samples
from the data stream into a matrix with eight rows
where each row represents one sensor. Each sample
can have a different length. We interpolate each sam-
ple to the maximum length of all samples.
6 DATA EXPLORATION
After collecting and pre-processing the data, we have
a visual look on it to see how each sensor contributes
to distinguishing users.
6.1 Comparing Time Series of the Same
User
In the first visualization, we compare all samples of
the same user by plotting all samples in the same plot.
We create one plot for each sensor, thus, resulting in
eight single time series plots. The plots for one of our
25 users are shown in Figure 3. In general, the plots
for the other users look similar and with at most two
outliers. We can see that the capacitive sensors (on
the left) seem to be more characteristic and expressive
Figure 3: This figure shows all door opening samples of one
user for each sensor: Bottom (Bo), Top (To), Back (Ba) and
Front (Fr) with their capacitive (C) and resistive (R) values,
respectively.
than the resistive ones. Another point that can be seen
is that all the samples in the plot have the same struc-
ture which shows that the door opening is not random
and follows a pattern.
6.2 Comparing the Time Series of
Different Users
In a second visualization, we compare the time se-
ries of two different users. We take two samples of
one user and one from another user and show one
of the plots in Figure 4. We see that the patterns of
user1 and user2 are similar and differ in the ampli-
tude. For the back (BaC) and front (FrC) capacitive
sensor, the sample of user2 is between the samples of
user1. On the other side, the sensors for bottom and
top (BoC and ToC) show a clear visible separation of
user1 and user2. This indicates that some sensors are
better suited for distinguishing the users than others.
Again, the plots of the other users show the same re-
sults with some plots showing a clear separation be-
tween the users in all eight plots.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
568
Figure 4: This figure shows the door opening from two sam-
ples of one user and one sample of a second participant for
each sensor: Bottom (Bo), Top (To), Back (Ba) and Front
(Fr) with their capacitive (C) and resistive (R) values, re-
spectively.
7 IDENTIFICATION RESULTS
In this work, we will do only a very first evalua-
tion of our data by analyzing the identification per-
formance with a closed-world assumption. Differ-
ent groupings of sensors are evaluated using five-
fold cross-validation and the common classifiers: K-
NN (k-nearest neighbors), SVM (support vector ma-
chine), Random Forests, Ada Boost and MLP (multi-
layer perceptron).
7.1 Evaluation per Sensor
For the first evaluation, we applied all classifiers to
each sensor individually to see how good each sen-
sor is to distinguish the users. Figure 5 and Figure 6
show the results (the mean accuracy of the five-fold
cross-validations runs) for each of the resistive and
capacitive sensors, respectively.
The plots show that the performance of the resis-
tive Sensors is significantly lower than the capacitive
sensors (best: 34.1% vs 74.6%) which correlate to the
Figure 5: The evaluation results for each of the resistive sen-
sors. The best accuracy is 34% from the top sensor (ToR)
using the random forest classifier.
Figure 6: The evaluation results for each of the capacitive
sensor. The best accuracy is 74% using from the front sen-
sor (FrC) using the random forest classifier.
data exploration observation. The top sensor (ToR)
performs best for all resistive sensors while it is differ-
ent for the capacitive sensors. The best result is given
by the front Sensor (FrC). In all cases, the random
forest classifier gives the best results on our dataset.
7.2 Evaluation of Sensor Groups
In the second evaluation, we grouped the sensors un-
der three categories: resistive only, capacitive only
and all sensors. Figure 7 shows the mean accuracies
of the cross-validations.
The combination of all resistive sensors improves
the identification result up to 57%. That’s better
but still inferior to one capacitive sensor. Grouping
all capacitive sensors achieved an accuracy of 88.6%
Again, the best overall classifier is the random forest.
The combination of all sensors reaches an accuracy
of 88.3% which gives a similar result as the capaci-
tive sensors only.
One reason why capacitive sensors perform better
Under Pressure: Pushing Down on Me – Touch Sensitive Door Handle to Identify Users at Room Entry
569
Figure 7: This Figures shows the accuracies of the evalua-
tion results of combined resistive sensors, combined capac-
itive and the combination of all sensors.
than resistive is that they have a higher resolution than
the resistive ones ([0, 2
14
] vs [0, 2
12
]) and a longer time
of interaction. This gives more significant data points
per sample and, therefore, are much more discrimina-
tive.
In our current identification procedure, the data of
the resistive sensors do not provide any information to
increase the identification result and therefore could
be ignored to increase the computation speed. How-
ever, the data is needed for other tasks, e.g. detecting
a door handle interaction (see Section 5).
Overall, we summarize that we can use touch-
sensitive door handles to identify users with a good
precision of over 88%.
7.3 Qualitative Evaluation
The evaluation of the questionnaire gives insights
about the participant’s opinion to the door handle au-
thentication system.
First, we look into the answers to the question of
how comfortable the usage of our prototype is. The
participants answered this question using a 1 to 5
scale where five means the prototype was indistin-
guishable from using a normal door handle. None
of the participants reported difficulties in using the
system and they evaluate the comfort with a mean of
4.54.
In a second question, we asked them whether
they think if such a door handle technology is an ac-
ceptable method for authentication. The participants
could answer with yes or no question to this question.
60% of the participants answered with yes. Besides,
the users could also add a reason for their answer.
They think that a smart door handle is easier and faster
to use than using a physical key and can provide more
security because it can not be easily stolen. On the
other hand, some participants are not convinced that
such a method can provide more security. They have
concerns about the precision and the uniqueness of
the door handle behavior.
8 CONCLUSION AND FUTURE
WORK
In this paper, we evaluate a system for identifying
users at the door opening using their touch behav-
ior. By building a prototype and a user study, we
recorded the capacitive and resistive data of 25 par-
ticipant’s door handle behavior. Data exploration and
classification showed that the capacitive sensors are
more suited for identifying users with an accuracy of
88% and using the random forest classifier. The ma-
jority of the participants agreed that this could be an
acceptable authentication method but also mentioned
concerns about the precision and uniqueness of the
door handle behavior.
The next steps are to extend the user study over
multiple days to analyze the robustness over the time
of this method as well as analyzing the influences of
different hands. For example, does the door open-
ing behavior change when we use the other hand, etc.
We will also add an accelerometer to the door han-
dle to analyze the door’s opening and closing move-
ment. Another step is to analyze the single phases of
the door opening process such as pushing down the
door handle, pushing or pulling open or closing the
door, etc. to improve the user authentication at the
door opening.
REFERENCES
A. Kumar, D. C. M. Wong, H. C. S. and Jain, A. K. (2003).
Personal Verification Using Palmprint and Hand Ge-
ometry Biometric. In Kittler, J. and Nixon, M. S., ed-
itors, Audio- and Video-Based Biometric Person Au-
thentication, pages 668–678.
Alam, S. and Yeasin, M. (2019). Person identification with
visual summary for a safe access to a smart home.
CoRR.
Aoyama, S., Ito, K., and Aoki, T. (2013). A multi-finger
knuckle recognition system for door handle. In 2013
IEEE Sixth International Conference on Biometrics:
Theory, Applications and Systems (BTAS), pages 1–7.
IEEE.
Bigun, J., Fierrez-Aguilar, J., Ortega-Garcia, J., and
Gonzales-Rodriguez, J. (2005). Combining biomet-
ric evidence for person authentication. In Advanced
Studies in Biometrics, pages 1–18. Springer.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
570
Brunelli, R. and Falavigna, D. (1995). Person identification
using multiple cues. In IEEE Transactions on Pattern
Analysis and Machine Intelligence, volume 17, pages
955–966.
ChaosComputerClub (2013). Chaos computer club breaks
apple touchid. [Online; posted 30-07-2019].
ChaosComputerClub (2017). Chaos computer clubs breaks
iris recognition system of the samsung galaxy s8. [On-
line; posted 30-July-2019].
Garcia, F. T., Krombholz, K., Mayer, R., and Weippl, E.
(2016). Hand dynamics for behavioral user authen-
tication. In 2016 11th International Conference on
Availability, Reliability and Security (ARES), pages
389–398. IEEE.
Greenberg, A. (2018). Hackers can steal a tesla model s in
seconds by cloning its key fob. [Online; posted 09-
October-2018].
Ishida, A., Murao, K., Terada, T., and Tsukamoto, M.
(2017). A user identification method based on fea-
tures of opening/closing a refrigerator door. In 2017
IEEE International Conference on Pervasive Comput-
ing and Communications Workshops (PerCom Work-
shops), pages 533–538. IEEE.
Kusanagi, D., Aoyama, S., Ito, K., and Aoki, T. (2017). A
practical person authentication system using second
minor finger knuckles for door security. IPSJ Trans-
actions on Computer Vision and Applications, 9(1):8.
Marsh, G., Borer, K., Salzbank, Z., Kim, J., and Pruitt, P.
(2014). Systems and methods for duplicating keys.
US Patent 8,682,468.
Mecke, L., Pfeuffer, K., Prange, S., and Alt, F. (2018). Open
sesame! user perception of physical, biometric, and
behavioural authentication concepts to open doors. In
Proceedings of the 17th International Conference on
Mobile and Ubiquitous Multimedia, MUM 2018, page
153–159, New York, NY, USA. Association for Com-
puting Machinery.
Odiete, J., Agbeyangi, A., and Olatinwo, O. (2017). An au-
tomated door control system using biometric technol-
ogy. IOSR Journal of Computer Engineering (IOSR-
JCE) 2278-0661, 19:20–25.
Pinkert, R. and Tanriverdi, H. (2018). Wie hacker in hotelz-
immer eindringen koennen. [Online; posted 25-April-
2018].
Schmidt, D., Chong, M. K., and Gellersen, H. (2010).
Handsdown: hand-contour-based user identification
for interactive surfaces. In Proceedings of the 6th
Nordic Conference on Human-Computer Interaction:
Extending Boundaries, pages 432–441.
Varasundar, M. and Balu, R. (2015). Web-based online em-
bedded door access control and home security system
based on face recognition. In International Confer-
ence on Circuits, Power and Computing Technologies.
Wahyudi, W. A. and Syazilawati, M. (2007). Intelli-
gent voice-based door access control system using
adaptive-network-based fuzzy inference systems (an-
fis) for building security. Journal of Computer Sci-
ence, 3:274–280.
Wendt, J. (2015). Apps ersetzen den schluesseldienst. [On-
line; posted 5-Januar-2015].
Under Pressure: Pushing Down on Me – Touch Sensitive Door Handle to Identify Users at Room Entry
571
