Fingerprinting Bluetooth-Low-Energy Devices Based on the Generic Attribute Profile

HAL Id: hal-02359914
https://hal.inria.fr/hal-02359914
Submitted on 16 Dec 2019
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entific research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Fingerprinting Bluetooth-Low-Energy Devices Based on
the Generic Attribute Profile
Guillaume Celosia, Mathieu Cunche
To cite this version:
Guillaume Celosia, Mathieu Cunche. Fingerprinting Bluetooth-Low-Energy Devices Based on the
Generic Attribute Profile. IoT S&P 2019 - 2nd International ACM Workshop on Security and Privacy
for the Internet-of-Things, Nov 2019, London, United Kingdom. pp.24-31, �10.1145/3338507.3358617�.
�hal-02359914�

Fingerprinting Bluetooth-Low-Energy Devices Based on the
Generic Attribute Profile
Guillaume Celosia
Univ Lyon, INSA Lyon, Inria, CITI
F-69621 Villeurbanne, France
guillaume.celosia@insa-lyon.fr
Mathieu Cunche
Univ Lyon, INSA Lyon, Inria, CITI
F-69621 Villeurbanne, France
mathieu.cunche@insa-lyon.fr
ABSTRACT
Bluetooth Low Energy (BLE) is a short range wireless technology in-
cluded in many consumer devices such as smartphones, earphones
and wristbands. As part of the Attribute (ATT) protocol, discover-
able BLE devices expose a data structure called Generic Attribute
(GATT) profile that describes supported features using concepts
of services and characteristics. This profile can be accessed by any
device in range and can expose users to privacy issues.
In this paper, we discuss how the GATT profile can be used to cre-
ate a fingerprint that can be exploited to circumvent anti-tracking
features of the BLE standard (i.e. MAC address randomization).
Leveraging a dataset of more than 13000 profiles, we analyze the
potential of this fingerprint and show that it can be used to uniquely
identify a number of devices. We also shed light on several issues
where GATT profiles can be mined to infer sensitive information
that can impact privacy of users. Finally, we suggest solutions to
mitigate those issues.
CCS CONCEPTS
• Networks → Network privacy and anonymity; • Security
and privacy → Mobile and wireless security.
KEYWORDS
Privacy; Bluetooth Low Energy; Tracking; Address randomization.
1 INTRODUCTION
Bluetooth is a widespread radio communication standard operat-
ing on the 2.4GHz ISM band. Bluetooth Low Energy (BLE) was
introduced [21, Vol 6] in 2010 by the Bluetooth Special Interest
Group (SIG) as a variant targeted towards battery-powered Internet
of Things (IoT) applications such as fitness trackers, headphones
and smartwatches. According to the Bluetooth SIG, more than two
billion devices supporting BLE have been shipped in 2018 [25].
To protect users against tracking [14], the Bluetooth Core Spec-
ification version 4.0 introduced the LE Privacy feature [21, Vol 3,
Part C, sec. 10.7] that defines the use of temporary and random
link layer identifiers. However, several works [4, 6] have discov-
ered flaws in its implementation showing that devices can still be
tracked despite LE Privacy provisions. BLE has also been subjected
to attacks aiming to infer sensitive information on users [7, 9].
In this paper, we focus on the Generic Attribute (GATT) pro-
file exposed by connectable BLE devices as part of the mandatory
Attribute (ATT) protocol. This profile presents a description of
features supported by a device through concepts of services and
characteristics. Moreover, as most of its elements are readable with-
out authentication, a GATT profile can be easily collected by any
device in range. We show that an attacker can use the content of
a GATT profile to compromise the privacy of the device owner
through tracking and inference of sensitive information.
Our contributions are outlined as follows:
• Based on a dataset of more than 13000 profiles, we demon-
strate that the content of a GATT profile can be leveraged to
build a fingerprint that can be used to single-out the device
and undermine the LE Privacy provisions (Section 5);
• We identify that some services and characteristics can be ex-
ploited to infer sensitive information on the user (Section 6);
• We provide a set of recommendations to mitigate the issues
uncovered in this work (Section 7).
Finally, we discuss related work in Section 8, and give concluding
remarks in Section 9.
2 BACKGROUND
2.1 BLE protocol
BLE is a radio communication standard [24] operating on the
2.4GHz ISM band. In BLE, devices can endorse two main roles:
Central and Peripheral. A Peripheral can be connectable or not
depending on if it accepts or not incoming connection requests. A
Central device can connect to a connectable Peripheral to communi-
cate with it. For instance, a smartphone can connect to a smartwatch
to send notifications and collect sensor readings.
BLE features a discovery mechanism that allows Central devices
to discover nearby Peripheral. As part of this mechanism, con-
nectable Peripheral periodically broadcast advertisement packets to
announce their presence.
2.2 BLE addressing and privacy
BLE devices are identified by a Bluetooth device address, a 48-bit
identifier that can be found within the payload of advertisement
packets. As part of its privacy feature (called LE Privacy), BLE has
introduced random addresses in addition to the globally unique
MAC address [1, sec. 8.2]. Thus, there are 4 types of device address in
BLE: Public, Random Static, Random Non-resolvable and Random
Resolvable. The Public device address corresponds to the MAC
address uniquely allocated to the device by the manufacturer. Other
device address types are acting as pseudonyms as they are randomly
generated and can change during the lifetime of a device.
Based on their temporal persistence, we classify those device
address types into two categories:
• Stable addresses: device addresses that are used by a device
indefinitely or for an extended period of time (i.e. Public and
Random Static addresses);

IoT S&P’19, November 15, 2019, London, United Kingdom G. Celosia and M. Cunche
• Private addresses: device addresses that are supposed to
change frequently1 (i.e. Random Non-resolvable and Random
Resolvable addresses).
2.3 GATT
In BLE, the Attribute (ATT) protocol is a Client/Server stateless
protocol based on attributes where devices can endorse each role
regardless of their BLE role (Peripheral or Central).
Data exposed by a Server are presented in a GATT profile which
is a hierarchical structure of attributes allowing the transfer of in-
formation between a Client and a Server. Within a GATT profile,
attributes can be either services or characteristics and are identified
by a universally unique identifier (UUID). In the hierarchical struc-
ture, conceptually related characteristics are grouped below a same
service (see Figure 1).
In addition to their UUID, characteristics are made up of an at-
tribute handle, a set of properties and a value. The handle specifies
the position of the characteristic in the profile while the value
holds the actual data of the characteristic. Properties are metadata
that specify which ATT operations (read, write, etc.) can be exe-
cuted on each particular attribute and with which specific security
requirements (encryption, authentication).
A service is identified by its UUID and is associated with two
handles (Handle Start and Handle End) that specify a range of char-
acteristics that are hierarchically dependent from this service.
Vendors are free to define their own services and character-
istics, but the Bluetooth SIG has already defined a number of
them [26]. For instance, the Bluetooth SIG has defined the Device
Information service that contains the Model Number String,
Software Revision String and System ID characteristics.
The BLE protocol features security mechanisms such as encryp-
tion and authentication that can be used in the ATT protocol. As
such, certain values of characteristics may only be accessed by an
authenticated Client [24, Vol 3, Part C, sec. 10.3]. Note that, the
value is the only element protected by this feature; the list of ser-
vices and characteristics as well as the associated metadata do not
require authentication to be accessed.
3 METHODOLOGY
Our study is based on a dataset of BLE GATT profiles from 13295
distinct Bluetooth device addresses collected over 5 months by the
authors during commute, work and leisure times. This dataset, pre-
sented in Table 2 and Table 3, was divided into two parts depending
on the nature of the address used by devices (Stable or Private).
Devices using Private addresses may be observed several times un-
der a different pseudonym. Thus, in the Private part of the dataset,
the number of actual devices is expected to be smaller than the
reported number of distinct device addresses.
This dataset was collected using a Raspberry Pi 3 single-board
computer equipped with four CSR v4.0 Bluetooth USB dongles. One
of those dongles continuously scans for advertising Peripheral using
the bluepy [13] python library. The 3 remaining dongles try to
connect to discovered connectable Peripheral prior to enumerate
attributes of their GATT profiles using our custom multi-threaded
1The Bluetooth Core Specification [24, Vol 3, Part C, App. A] recommends to renew
Random Non-resolvable and Random Resolvable addresses at most every 15 minutes.
GATT Proﬁle
UUID Properties Value
Handle Start
Handle End
Handle Start
Handle End
Handle
Handle
Handle
UUID
UUID
Service Characteristics
Service Characteristics
Figure 1: Structure of a GATT profile. Services are composed
of a UUID along with two handles (Handle Start and Handle
End) delimiting the hierarchically dependent characteristics.
Characteristics are each constituted of a handle, a UUID, a
set of properties and a value containing data.
version of the bleah [15] python tool. Each GATT profile is then
structured as a json string and stored for further analysis.
Dataset anonymization: To limit the privacy risks associated
with the collected data, we applied modifications to prevent user re-
identification. We focused on attributes corresponding to identifiers
(Stable device addresses, device names, etc.) and on temporal data
(timestamps). The 24 least significant bits (the NIC) of Stable ad-
dresses have been pseudonymized through keyed-hashing2. String
identifiers potentially containing names of users were sanitized by
searching and key-hashing substrings that were matching names.
Finally, the temporal information has been transformed from abso-
lute (date and time) to relative (time elapsed since the beginning of
the collection campaign).
4 ATTACKER MODEL
We consider an active attacker which monitors the BLE advertising
channels to detect nearby connectable Peripheral, connect to them
and collect their GATT profiles using several ATT Read By Type Re-
quest. Furthermore, we assume that the device used by the attacker
has not been paired with any Peripheral: it cannot authenticate
itself and access protected values of characteristics. As described
in Section 3, those assumptions can be satisfied using off-the-shelf
hardware and open-source software. In addition, we found that
a full GATT profile can be collected in a matter of seconds (see
Table 4). On the target side, we assume that the device has its BLE
interface turned on, is in communication range and is discoverable.
Based on the collected GATT profile, the attacker can have two
objectives: 1) generate a fingerprint of the device in order to track
it despite its address randomization scheme and 2) infer sensitive
information on the device and its owner.
5 GATT PROFILES FINGERPRINTING
Following the approach of Vanhoef et al. [28], we study how much
identifying information can be found in GATT profiles. In particular,
we study how services and characteristics can be used to create a
fingerprint of the device. In case this fingerprint is unique enough,
it can be used to track a device despite the address randomization.
2The key used during this process has been erased.

Fingerprinting BLE Devices Based on the GATT Profile IoT S&P’19, November 15, 2019, London, United Kingdom
5.1 GATT fingerprint artifacts
The GATT profile of a BLE device is a data structure that can
be easily accessed and that includes a number of data elements
that can be used for fingerprinting. First, the number of possible
components is large: online GATT specifications [26] describe a
list of 40 services and 226 characteristics that can be complemented
by vendors with their own custom elements. In total, we found
263 distinct services and 1086 distinct characteristics in the dataset.
In addition, characteristics are associated with a value that can
contain up to 512 bytes of data [24, Vol 3, Part F, sec. 3.2.9]. All
those elements are also accompanied by metadata: handles and
properties respectively represented by 2 bytes and 8 flags.
Second, the content of a profile will vary depending on the device
type as the GATT profile reflects the features and the characteristics
of the device. For instance, the Cycling Power Measurement char-
acteristic will be only included in cycle devices, while only weight
scales will expose the Weight Scale Feature characteristic.
Finally, values associated with characteristics may vary from
one device to another as they can reflect the device state or identity.
For instance, this is the case of identifiers such as the Device Name
and Model Number String.
Overall, a GATT profile is a data structure containing a large
number of elements that are subject to variation between devices,
and thus hold a potential for fingerprinting.
To create the fingerprint of a BLE device, we considered the
following artifacts:
• List of services, including for each service:
– Handles (start-end): the handle range associated with
the service (two 16-bit identifiers)
– UUID: the UUID associated with the service (128-bit iden-
tifier)
• List of characteristics, including for each characteristic:
– Handle: the handle associated with the characteristic (16-
bit identifier)
– UUID: the UUID associated with the characteristic (128-
bit identifier)
– Properties: the properties of the characteristic (8 bits)
– Value: the value of the characteristic (from 0 to 512 bytes)
5.2 Fingerprinting evaluation
We use the collected dataset to evaluate the fingerprinting potential
offered by the GATT profiles. Previously presented artifacts were
extracted from the dataset and stored in a database in which each
fingerprint is associated with a device address and a timestamp.
Then, the resulting database was processed to compute fingerprint-
ing metrics: entropy (Section 5.3) and anonymity sets (Section 5.4).
Impact of random addresses on evaluation: The dataset in-
cludes records from devices using random addresses (Private ad-
dresses). A device using the address randomization scheme can be
observed multiple times under different pseudonyms and thus the
corresponding fingerprint will be counted multiple times instead
of one. This overcounting will have an impact on the privacy met-
rics: the entropy will be reduced and the size of the anonymity set
will be increased. Therefore, values reported for the Private part
of the dataset should be considered as an underestimation of the
fingerprinting potential.
5.3 Empirical entropy
Leveraging the dataset, we evaluate the quantity of information
brought by the services and characteristics. The entropy is a metric
used to measure the amount of identifying information brought
by an element of the fingerprint [8]. The database of fingerprints
was processed to compute an empirical evaluation of the entropy
of each artifact i using the following formula:
Hi = −
Õ
j ∈Ei
fi,j ∗ log2 fi,j (1)
where Ei is the domain of possible values for an artifact i and fi,j is
the frequency (i.e. probability) of the value j for the artifact i in the
dataset. Note that, the absence of an artifact was also considered as
a possible value.
Table 1 presents the entropy for the 8 most common services
and characteristics exposed in the dataset as well as for the overall
profile. The Entropy column presents the amount of identifying
bits provided by the artifacts. The Stability column presents the
fraction of devices observed several times for which the value of
the artifact is constant throughout the dataset. Finally, the Affected
devices column presents the fraction of devices that include this
artifact in their GATT profiles.
A first observation is the high stability of the fingerprint: the
overall fingerprint is stable in more than 95% of the cases. The
entropy of single artifacts is typically comprised between 0 and
2 bits. However, some artifacts such as the Device Name and the
Model Number String characteristic can bring up to 3.152 bits of
information. Indeed, those artifacts are in fact identifiers.
Variations can be also observed between the types of device
address: the Device Name brings less information for Private than
for Stable addresses. Actually, we observed that for devices using
Private addresses, this characteristic is often configured to carry a
generic value3. This is likely a deliberate choice done for privacy
reasons. However, developers appear to have overlooked the Model
Number String as it appears to be a high source of information for
Private addresses (2.757 bits).
Overall, characteristics appear to bring more information than
services (4.380 bits against 2.111 bits). This is explained by the fact
that characteristics hold more artifacts than services. When consid-
ering the full fingerprint, which includes both the characteristics
and the services, we can observe that the entropy is the same as
with the characteristics alone. This is due to the fact that artifacts
of a service (handles and UUID) are fully determined by artifacts
of its characteristics (remind that characteristics are hierarchically
dependent from services). In other words, services do not bring
additional information with regard to characteristics.
5.4 Anonymity sets
To further study the fingerprinting potential of GATT profiles, we
used the concept of anonymity set, which is defined as a set of
entities that share the same fingerprint. From a privacy point of
view, the larger the anonymity set the better.
Aided by the kmap [11] python tool, we computed the anonymity
sets for the fingerprints contained in the dataset. Figures 2a and 2b
3For instance, the value of the Device Name characteristic is iPhone for both an Apple
iPhone 6 and an Apple iPhone 8 smartphone.

Table 1: Empirical entropy computed from the dataset for services and characteristics exposed within GATT profiles. For each
item: the entropy brought by the attribute, the percentage of devices for which this item is stable over time, and the percentage
of devices that include this item in their GATT profiles.
Services
<Hdle
Start,
Hdle
End,
UUID>
Entropy (bits) Stability* (%) Affected devices (%)
All Stable Private All Stable Private All Stable addr. Private addr.
Generic Access 0.750 0.606 0.245 100 100 100 99.01 92.99 100
Generic Attribute 0.728 0.560 0.248 100 100 100 97.34 81.35 99.97
Apple Continuity 0.680 0.317 0.462 100 100 100 84.74 6.48 97.64
Apple Nearby Service 0.720 0.306 0.499 100 100 100 84.32 4.36 97.50
Device Information 1.425 0.551 1.037 100 100 100 69.74 55.90 72.02
Battery Service 0.943 0.328 0.879 100 100 100 57.86 5.58 66.48
Current Time Service 0.871 0.277 0.866 100 100 100 57.08 0.05 66.48
Apple Media Service 0.835 0.277 0.831 100 – 100 57.07 0 66.48
Overall 2.111 0.803 1.330 99.28 100 99.20 – – –
Characteristics
<Hdle,
UUID,
Prop,
Value>
Device Name 1.913 1.191 0.731 100 100 100 99.65 97.56 100
Appearance 1.148 0.625 0.578 100 100 100 98.90 92.40 99.97
Service Changed 0.766 0.566 0.290 100 100 100 97.34 81.35 99.97
Apple Continuity 0.680 0.317 0.462 100 100 100 84.74 6.48 97.64
Apple Nearby 0.720 0.306 0.499 100 100 100 84.32 4.36 97.50
Manuf. Name String 1.422 0.538 1.053 99.82 100 99.81 69.38 53.40 72.02
Model Number String 3.152 0.564 2.757 99.82 100 99.81 69.32 52.98 72.02
Battery Level 1.020 0.395 0.879 100 100 100 58.65 11.16 66.48
Overall 4.380 1.294 3.092 97.84 95.71 98.08 – – –
Overall (services + char-
acteristics)
4.380 1.294 3.092 97.84 95.71 98.08 – – –
* Stability values have been computed only from device addresses that we observed multiple times.
respectively show the distributions of the set sizes for Stable and
Private addresses.
For Stable addresses, the anonymity sets are small with 94.75%
of sets of size 1, meaning that those devices can be uniquely identi-
fied by their fingerprints. This is not critical, as those devices can
be already identified through their Stable addresses. However, it
demonstrates the potential for unique identification based on the
GATT profile.
Moving to Private addresses, we observe that a smaller number of
devices are uniquely identifiable (4.28%) and that 74.33% of devices
are in anonymity sets of size 100 or more. This improvement could
be explained by the fact that vendors have reduced the amount of
identifying information included in GATT profiles of devices using
Private addresses.
Focusing on devices using Private addresses, we found that a
large number of them are Apple iPhone smartphones4. By dividing
the Private part of the dataset between non-iPhone devices (Fig-
ure 2c) and iPhones (Figure 2d), we found that a majority of iPhones
were sharing their fingerprints with many other devices: 85.49% are
in anonymity sets of size 100 or more. On the other end, non-iPhone
devices using Private addresses have less common fingerprints as
74.38% of them are in anonymity sets of size 10 or less, and 32.09%
of them are unique.
A possible explanation to this phenomenon is that Apple dis-
tributes a large number of devices but focused on a small number
4This device model identification is based on the values of the Model Number String
and Manufacturer Name String characteristics along with the presence of Apple
specific services (Apple Continuity Service, Apple Nearby Service, etc.).
of models (a single line of products with few variants per gen-
eration). Furthermore, the software running on those devices is
homogeneous. It seems that a side effect of Apple commercial and
technical policies is to reduce possibilities of uniquely identifying
their devices based on technical characteristics.
6 INFERRING INFORMATION FROM GATT
PROFILES
The content of a GATT profile can be leveraged to infer information
on the device and its user. In this section, we present a number of
elements found in GATT profiles that can be used to infer poten-
tially sensitive information. In particular, we found that the value
of characteristics was a rich source of information and was often
readable without authentication (see Table 9).
We identified that information found in GATT profiles can be
used to infer the following information: device type, device model,
device manufacturer and user’s name. All this information can
threaten the privacy of the device owner. Information on the device
model can lead to inventory attacks [9], and user’s name can reveal
the identity of the owner. Furthermore, we found that the value of
some characteristics can hold identifiers, which can be leveraged
for tracking despite the device address randomization.
6.1 Human readable identifiers
GATT profiles can include characteristics for which the value is a
human readable string. For instance, this is the case of the Device
Name, Model Number String and Manufacturer Name String.
The Device Name characteristic is available and readable in more
than 99% of the profiles, and often includes names of manufacturer,

(a) Stable addr.
(1882 devices)
(b) Private addr.
(11413 devices)
(c) Private without iPhones
(1489 devices)
(d) Private iPhones only
(9924 devices)
Figure 2: Anonymity sets of GATT profiles in the dataset. The dot size is proportional to the number of devices in the set.
model and user such as Polar M400 and Alice’s MacBook Pro.
Similarly, values carried by the Model Number String explicitly
identify the device model [2]. For instance, the model number of an
Apple iPhone 8 is iPhone10,4 while iPad8,3 indicates a 11-inch
Apple iPad Pro. Finally, the Manufacturer Name String directly
reveals the manufacturer of the device.
6.2 Digital identifiers
Serial number strings: The Serial Number String character-
istic carries a variable-length utf-8 string representing the serial
number for a particular instance of the device. Leveraging the
dataset, we found that the format of this identifier is often specific
to a vendor5 and thus can be leveraged to infer the manufacturer.
System IDs: The value of the System ID characteristic is a 64-bit
structure which consists of a 40-bit manufacturer-defined identifier
concatenated to a 24-bit Organizationally Unique Identifier (OUI).
By definition, this OUI is issued by the IEEE Registration Authority
to companies and thus can reveal the manufacturer. For instance,
manufacturers such as Xiaomi and Amazfit include their OUI within
the value of their System ID characteristic.
PnP IDs: The PnP ID characteristic carries a set of values that
are used to create a unique identifier for the device. Included in
this characteristic is a Vendor ID Source, a Vendor ID, a Product
ID and a Product Version field. The Vendor ID Source specifies the
type of the Vendor ID value: a company identifier assigned by the
Bluetooth SIG [23] or a value assigned by the USB Implementers
Forum [10]. We observed that manufacturers such as Gigaset (com-
pany ID 0x0180) and Freebox (USB ID 0x10eb) include an identifi-
able PnP ID in their GATT profiles. Thus, the PnP ID can reveal the
manufacturer but also the device model and software version. For
instance, Product IDs of Fitbit Surge and Fitbit Charge are respec-
tively 0x0010 and 0x0013 while the Product Version of the Bose
SoundSport Free earphones is 0x0132 corresponding to the value
carried by its Software Revision String (1.3.2).
6.3 Enumerated type values
Some characteristics are associated with enumerated values whose
meanings are specified by the GATT specifications [26].
5For instance, Ultimate Ears and Bose respectively code their serial numbers following
the "ˆ1...LZ0....800$" and "ˆ07....[Z,P][6,7,8]" regular expressions.
Appearances: The Appearance characteristic represents informa-
tion about the external appearance of the device. Such a charac-
teristic is readable in more than 99% of devices and can provide a
broad description (e.g. Generic Tag) as well as a more specific one
(e.g. Running Walking Sensor: On-Shoe). For instance, appear-
ance values of Apple TV and Garmin Forerunner 230 devices are
respectively Generic Media Player and Watch: Sports Watch.
Moreover, as reported in [6], certain appearance values indicate
specific medical devices such as an Insulin Pen that could trivially
betray a medical condition of the owner.
Sensor locations: The Body Sensor Location and Sensor Loca-
tion characteristics are indicating where the device is located on
the user. As such, they provide clues to infer the type of device: the
presence of one of those characteristics indicates that the device is
a sensor, and its value can further specify the type of sensor (see
Table 5 and Table 6 for respectively a list of body sensor location
and sensor location values).
6.4 Measurement values
During our study, we discovered that Mio Global Alpha 2 smart-
watches expose a readable value within their Running Speed and
Cadence (RSC) Measurement characteristic revealing the physical
activity of the user (i.e. walking or running). As a result, such a
characteristic can constitute an additional source of information
that could be used to profile or physically identify a user [7].
6.5 Names of services and characteristics
Beyond values carried by characteristics, names of services and
characteristics can be leveraged as an indicator to reveal both the
device type and manufacturer. For instance, the Cycling Speed
and Cadence (CSC) Feature characteristic will be only included
in cycle devices, while the presence of the RSC Measurement char-
acteristic denotes running sensors. In addition, we found that UUID
of attributes exposed within GATT profiles can be customized by
manufacturers. Leveraging online specifications and codes such as
the Apple Notification Center Service [3] and Xiaomi Mi Band 2 [27]
ones, it becomes possible to uncover meanings of custom UUIDs
disclosing the corresponding manufacturer at the same time.

7 RECOMMENDATIONS
In light of the presented issues, we provide recommendations that
should be considered by manufacturers of BLE devices, but also by
the Bluetooth SIG to improve the Bluetooth Core Specification.
Restricting access to values of characteristics: We found that a
number of characteristics are readable for unauthenticated Clients
which expose the device to fingerprinting and inference of sensitive
information. In many cases, it is not clear why this information
is left openly available. A simple solution is to use the permission
system of GATT to ensure that those values can only be read by
authenticated Clients. This mechanism should be set by default
on all characteristics and its removal should be justified by a valid
requirement.
To support this statement, we leveraged the dataset to simulate
the adoption of this measure. All values of characteristics were
removed at the exception of mandatory ones (values of the Device
Name, Appearance and Service Changed characteristics) and the
entropy was re-evaluated (see Table 8). Removing those values has
a significant impact on the fingerprinting potential: the overall
entropy goes down from 4.380 to 2.765 bits.
In cases where this solution cannot be adopted, values should be
as general as possible. For instance, the value of the Model Number
String characteristic within Apple devices could be just Apple or
iOS instead of iPhone10,4 or iPad8,3 (as reported by [16]).
Minimizing exposure of GATT profile: The current version of
Bluetooth Core Specification [24, Vol 3, Part G, sec. 8.1] specifies
the following: "the list of services and characteristics that a device
supports is not considered private or confidential information, and
therefore the service and characteristic discovery procedures shall al-
ways be permitted". We demonstrated that this is not the case: even
if values are not readable, the list of services and characteristics
available in a GATT profile can be used for fingerprinting and in-
ference of sensitive information. A potential mitigation technique
would be to minimize the exposure of the GATT profile. This could
be done by setting access control properties to services and char-
acteristics so that only authenticated Clients can access them. By
default, an unauthenticated Client will only see basic services and
characteristics (e.g. Generic Attribute and Service Changed),
and only authenticated Clients will be able to see the full list.
8 RELATED WORK
The possibility of singling-out a device based on its technical char-
acteristics and attributes has been explored by several works. This
technique has been used in the context of Web browser in order
to track users despite anti-tracking mechanisms [8, 12, 20]. In the
context of wireless devices, fingerprinting has been used to iden-
tify technical characteristics of a device such as the version of the
operating system or driver [5], as well as the device model [16].
The timing of frame transmissions has been leveraged in 802.11
networks to identify a device [19] and to track it over time [18].
In [17, 28], the authors have demonstrated that the content of 802.11
probe request frames can be used to fingerprint devices and defeat
MAC address randomization. Our work shows that this problem is
not limited to 802.11 and that BLE suffers from similar issues.
Recent works [4, 6, 16] have studied the advertisement mecha-
nism of BLE and showed that the content of advertisement pack-
ets can be used to defeat address randomization. Our work shows
that address randomization is also threatened by the content of
GATT profiles.
Beyond tracking issues, BLE can be at the source of private
information leakages. Das et al. demonstrated [7] how the BLE
traffic from a fitness tracker can leak the physical activity of the
wearer. In [9], Fawaz et al. discussed how information exposed in
BLE advertisement packets can be leveraged as side information to
infer sensitive attributes such as a medical condition. In this paper,
we continue this research direction by presenting a detailed list of
information leakages affecting BLE devices, that we illustrate with
examples extracted from a large real-world dataset.
9 CONCLUSION
In this work, we studied data exposed within GATT profiles of BLE
devices. First, we demonstrated that the content of a GATT pro-
file can be leveraged to fingerprint a device. Indeed, identifiers and
values composing this profile are diverse enough to act as a finger-
print. This fingerprint can be used to track a device even if it uses
anti-tracking mechanisms such as the MAC address randomization.
This contribution complements recent works [4, 6, 16, 17, 28] that
have demonstrated the difficulties in the implementation of device
address randomization.
Then, we showed how the data exposed in GATT profiles can
be mined to infer information on the device such as its type, model,
manufacturer and software version, which can be leveraged to
threaten privacy of users.
To the best of our knowledge, there is no indication that those
techniques are currently exploited in the wild. In the Web ecosys-
tem, introduction of anti-tracking techniques has triggered [20]
the deployment of fingerprinting techniques by Web trackers. The
recent adoption of address randomization in wireless technolo-
gies [16, 28] could trigger a similar move by the industry of physical
tracking. As a consequence, it is important to take this threat into
account as soon as possible, by reviewing and complementing the
Bluetooth Core Specification with additional requirements.
ACKNOWLEDGEMENTS
This work was supported by the SPIE ICS-INSA Lyon IoT chair
and the SPARTA project. The authors would like to thanks Alexis
Duque for his valuable insight on BLE.
REFERENCES
[1] 2014. IEEE Standard for Local and Metropolitan Area Networks: Overview and
Architecture. IEEE Std 802-2014 (Revision to IEEE Std 802-2001) (June 2014), 1–74.
https://doi.org/10.1109/IEEESTD.2014.6847097
[2] 2019. Models - The iPhone Wiki. (2019). https://www.theiphonewiki.com/wiki/
Models Accessed: 2019-07-04.
[3] Apple. 2014. Apple Notification Center Service (ANCS) Specification. (2014).
https://developer.apple.com/library/archive/documentation/CoreBluetooth/
Reference/AppleNotificationCenterServiceSpecification/Specification/
Specification.html#//apple_ref/doc/uid/TP40013460-CH1-SW8 Accessed:
2019-07-04.
[4] Johannes K Becker, David Li, and David Starobinski. 2019. Tracking Anonymized
Bluetooth Devices. Proceedings on Privacy Enhancing Technologies 2019, 3 (2019),
50–65.
[5] Sergey Bratus, Cory Cornelius, David Kotz, and Daniel Peebles. 2008. Active
Behavioral Fingerprinting of Wireless Devices. In Proceedings of the First ACM

Conference on Wireless Network Security (WiSec ’08). ACM, New York, NY, USA,
56–61. https://doi.org/10.1145/1352533.1352543
[6] Guillaume Celosia and Mathieu Cunche. 2019. Saving Private Addresses: An
Analysis of Privacy Issues in the Bluetooth-Low-Energy Advertising Mechanism.
In Proceedings of the 16th EAI International Conference on Mobile and Ubiquitous
Systems: Computing, Networking and Services. ACM.
[7] Aveek K Das, Parth H Pathak, Chen-Nee Chuah, and Prasant Mohapatra. 2016.
Uncovering privacy leakage in BLE network traffic of wearable fitness trackers.
In Proceedings of the 17th International Workshop on Mobile Computing Systems
and Applications. ACM, 99–104.
[8] Peter Eckersley. 2010. How unique is your web browser ?. In International
Symposium on Privacy Enhancing Technologies Symposium. Springer, 1–18.
[9] Kassem Fawaz, Kyu-Han Kim, and Kang G Shin. 2016. Protecting Privacy of BLE
Device Users. In USENIX Security Symposium. 1205–1221.
[10] USB Implementers Forum. 2019. Membership Lookup & List. (2019). https:
//www.usb.org/members Accessed: 2019-07-04.
[11] Gabor Gyorgy Gulyas, Gergely Acs, and Claude Castelluccia. 2016. Near-optimal
fingerprinting with constraints. Proceedings on Privacy Enhancing Technologies
2016, 4 (2016), 470–487.
[12] Gabor Gyorgy Gulyas, Doliere Francis Some, Nataliia Bielova, and Claude Castel-
luccia. 2018. To Extend or Not to Extend: On the Uniqueness of Browser Ex-
tensions and Web Logins. In Proceedings of the 2018 Workshop on Privacy in
the Electronic Society (WPES’18). ACM, New York, NY, USA, 14–27. https:
//doi.org/10.1145/3267323.3268959
[13] Ian Harvey. 2014. bluepy - a Bluetooth LE interface for Python. https://ianharvey.
github.io/bluepy-doc/index.html Accessed: 2019-07-04.
[14] Markus Jakobsson and Susanne Wetzel. 2001. Security Weaknesses in Bluetooth.
In Topics in Cryptology - CT-RSA 2001 (Lecture Notes in Computer Science), David
Naccache (Ed.). Springer Berlin Heidelberg, 176–191.
[15] Simone Margaritelli. 2017. This Is Not a Post About BLE, Introducing
BLEAH. https://www.evilsocket.net/2017/09/23/This-is-not-a-post-about-BLE-
introducing-BLEAH/ Accessed: 2019-07-04.
[16] Jeremy Martin, Douglas Alpuche, Kristina Bodeman, Lamont Brown, Ellis Fenske,
Lucas Foppe, Travis Mayberry, Erik Rye, Brandon Sipes, and Sam Teplov. 2019.
Handoff All Your Privacy – A Review of Apple’s Bluetooth Low Energy Continuity
Protocol. Proceedings on Privacy Enhancing Technologies 2019, 4 (2019), 34–53.
[17] Jeremy Martin, Travis Mayberry, Collin Donahue, Lucas Foppe, Lamont Brown,
Chadwick Riggins, Erik C Rye, and Dane Brown. 2017. A study of MAC ad-
dress randomization in mobile devices and when it fails. Proceedings on Privacy
Enhancing Technologies 2017, 4 (2017), 365–383.
[18] Celestin Matte, Mathieu Cunche, Franck Rousseau, and Mathy Vanhoef. 2016.
Defeating MAC Address Randomization Through Timing Attacks. In Proceedings
of the 9th ACM Conference on Security & Privacy in Wireless and Mobile Networks
(WiSec ’16). ACM, New York, NY, USA, 15–20. https://doi.org/10.1145/2939918.
2939930 event-place: Darmstadt, Germany.
[19] Christoph Neumann, Olivier Heen, and Stephane Onno. 2012. An empirical study
of passive 802.11 device fingerprinting. In 2012 32nd International Conference on
Distributed Computing Systems Workshops. IEEE, 593–602.
[20] N. Nikiforakis, A. Kapravelos, W. Joosen, C. Kruegel, F. Piessens, and G. Vi-
gna. 2013. Cookieless Monster: Exploring the Ecosystem of Web-Based De-
vice Fingerprinting. In 2013 IEEE Symposium on Security and Privacy. 541–555.
https://doi.org/10.1109/SP.2013.43
[21] Bluetooth SIG. 2010. Bluetooth Core Specification v4.0. https://www.bluetooth.
org/docman/handlers/downloaddoc.ashx?doc_id=456433 Accessed: 2019-07-04.
[22] Bluetooth SIG. 2017. GATT Specification Supplement v1.0. https://www.bluetooth.
[23] Bluetooth SIG. 2019. Assigned Numbers - Company Identifiers. (2019). https://
www.bluetooth.com/specifications/assigned-numbers/company-identifiers/ Ac-
cessed: 2019-07-04.
[24] Bluetooth SIG. 2019. Bluetooth Core Specification v5.1. https://www.bluetooth.
[25] Bluetooth SIG. 2019. Bluetooth Market Update 2019. Technical Re-
port. https://www.bluetooth.com/wp-content/uploads/2018/04/2019-Bluetooth-
Market-Update.pdf Accessed: 2019-07-04.
[26] Bluetooth SIG. 2019. GATT Specifications. (2019). https://www.bluetooth.com/
specifications/gatt/ Accessed: 2019-07-04.
[27] Leo Soares. 2019. MiBand2 - Python library to work with Xiaomi MiBand 2.
(2019). https://github.com/creotiv/MiBand2/blob/master/constants.py Accessed:
2019-07-04.
[28] Mathy Vanhoef, Celestin Matte, Mathieu Cunche, Leonardo S. Cardoso, and Frank
Piessens. 2016. Why MAC Address Randomization is Not Enough: An Analysis
of Wi-Fi Network Discovery Mechanisms. In Proceedings of the 11th ACM on
Asia Conference on Computer and Communications Security (ASIA CCS ’16). ACM,
New York, NY, USA, 413–424. https://doi.org/10.1145/2897845.2897883
A APPENDIX
Table 2: Composition of the dataset of GATT profiles for
each type of device address.
Stable Private
Public Static Non-res. Res.
Device addr.
% 10.29 3.87 0.17 85.67
# 1368 514 23 11390
Table 3: Top 10 services and characteristics in the dataset.
Device addr. (%)
Services
All Stable Private
Generic Access 99.01 92.99 100
Generic Attribute 97.34 81.35 99.97
Apple Continuity 84.74 6.48 97.64
Apple Nearby Service 84.32 4.36 97.50
Device Information 69.74 55.90 72.02
Battery Service 57.86 5.58 66.48
Current Time Service 57.08 0.05 66.48
Apple Media Service 57.07 0 66.48
Apple NCS Service 57.07 0 66.48
ISSC Transparent 3.86 27.26 0
Overall* 99.17 94.16 100
Characteristics
Device Name 99.65 97.56 100
Appearance 98.90 92.40 99.97
Service Changed 97.34 81.35 99.97
Apple Nearby 84.32 4.36 97.50
Manuf. Name String 69.38 53.40 72.02
Model Number String 69.32 52.98 72.02
Battery Level 58.65 11.16 66.48
Current Time 57.30 1.65 66.48
Apple MS Entity Att. 57.07 0 66.48
Overall* 99.96 99.73 100
* Devices that include at least one of the top 10 service or characteristic.

Table 4: Average time to collect a GATT profile among dif-
ferent devices.
Device type Device
Time
(sec)
Lightbulb Osram Smart+ 6.531
Motion sensor Eve Motion 6.468
Socket outlet Eve Energy 5.919
Smartphone Apple iPhone 8 4.354
Smartphone Apple iPhone 6 4.259
Keyring Nut 4.148
TV dongle Google Chromecast 3.660
Fitness wristband Fitbit Inspire 3.231
Presentation remote Logitech Spotlight 2.860
Smartwatch Apple Watch Series 3 2.853
Heart rate monitor Polar H7 2.751
Fitness wristband Fitbit Flex 2.552
Headset Bose SoundLink Around-Ear II 2.181
Speaker Divacore Ktulu2+ 1.742
Keyring Chipolo 1.426
Average 3.662
Table 5: List of Body Sensor Location values extracted
from [22, sec. 3.24.2.1].
Value Description
0x01 Chest
0x02 Wrist
0x03 Finger
0x04 Hand
0x05 Ear Lobe
0x06 Foot
Table 6: List of Sensor Location values extracted from [22,
sec. 3.148.2.1].
Value Description
0x00 Other
0x01 Top of shoe
0x02 In shoe
0x03 Hip
0x04 Front Wheel
0x05 Left Crank
0x06 Right Crank
0x07 Left Pedal
0x08 Right Pedal
0x09 Front Hub
0x0a Rear Dropout
0x0b Chainstay
0x0c Rear Wheel
0x0d Rear Hub
0x0e Chest
0x0f Spider
0x10 Chain Ring

Table 7: Example of a GATT profile collected from an Apple iPhone 8 smartphone.
Handles Service > Characteristics (UUID) Properties Value
0001 -> 0005 Generic Access (00001800-0000-1000-8000-00805f9b34fb)
0003 Device Name (00002a00-0000-1000-8000-00805f9b34fb) Read iPhone
0005 Appearance (00002a01-0000-1000-8000-00805f9b34fb) Read Generic Phone
0006 -> 0009 Generic Attribute (00001801-0000-1000-8000-00805f9b34fb)
0008 Service Changed (00002a05-0000-1000-8000-00805f9b34fb) Indicate –
000a -> 000e Apple Continuity Service (d0611e78-bbb4-4591-a5f8-487910ae4366)
000c Continuity Characteristic (8667556c-9a37-4c91-84ed-54ee27d90049) Notify, Write, Extended properties –
000f -> 0013 Apple Nearby Service (9fa480e0-4967-4542-9390-d343dc5d04ae)
00011 Nearby Characteristic (af0badb1-5b99-43cd-917a-a77bc549e3cc) Notify, Write, Extended properties –
0014 -> 0017 Battery Service (0000180f-0000-1000-8000-00805f9b34fb)
00016 Battery Level (00002a19-0000-1000-8000-00805f9b34fb) Notify, Read –
0018 -> 001d Current Time Service (00001805-0000-1000-8000-00805f9b34fb)
001a Current Time (00002a2b-0000-1000-8000-00805f9b34fb) Notify, Read –
001d Local Time Information (00002a0f-0000-1000-8000-00805f9b34fb) Read –
001e -> 0022 Device Information (0000180a-0000-1000-8000-00805f9b34fb)
0020 Manufacturer Name String (00002a29-0000-1000-8000-00805f9b34fb) Read Apple Inc.
0022 Model Number String (00002a24-0000-1000-8000-00805f9b34fb) Read iPhone10,4
0023 -> 002c Apple Notification Center Service (7905f431-b5ce-4e99-a40f-4b1e122d00d0)
0025 Control Point (69d1d8f3-45e1-49a8-9821-9bbdfdaad9d9) Write, Extended properties –
0028 Notification Source (9fbf120d-6301-42d9-8c58-25e699a21dbd) Notify –
002b Data Source (22eac6e9-24d6-4bb5-be44-b36ace7c7bfb) Notify –
002d -> 0038 Apple Media Service (89d3502b-0f36-433a-8ef4-c502ad55f8dc)
002f Remote Command (9b3c81d8-57b1-4a8a-b8df-0e56f7ca51c2) Notify, Write, Extended properties –
0033 Entity Update (2f7cabce-808d-411f-9a0c-bb92ba96c102) Notify, Write, Extended properties –
0037 Entity Attribute (c6b2f38c-23ab-46d8-a6ab-a3a870bbd5d7) Read, Write, Extended properties –
Table 8: Empirical entropy of characteristics exposed within GATT profiles in the dataset, without the values (only handles,
UUIDs and properties are considered).
Entropy (bits)
All Stable addr. Private addr.
Characteristics
<Hdle,
UUID,
Prop>
Device Name (+ value*) 1.913 1.191 0.731
Appearance (+ value*) 1.148 0.625 0.578
Service Changed (+ value*) 0.766 0.566 0.290
Apple Nearby 0.720 0.306 0.499
Manuf. Name String 1.372 0.517 1.031
Model Number String 1.327 0.491 1.030
Battery Level 0.985 0.361 0.879
Overall 2.764 1.208 1.564
Overall (services + characteristics) 2.765 1.208 1.564
* Value of this characteristic is kept during the entropy computation because the Bluetooth Core Specifica-
tion specifies it as mandatory.

Table 9: List of Bluetooth SIG-defined characteristics values that have been observed as readable at least once in the dataset.
Percentages reported for readable values are fractions of devices for which this value was readable.
Readable values
All Stable Private
% # % # % #
Device Name 99.49 13182 99.18 1821 99.54 11361
Appearance 99.48 13082 98.56 1714 99.62 11368
Service Changed 0.02 2 0.13 2 0 0
Manufacturer Name String 99.48 9177 99.40 999 99.49 8178
Model Number String 99.36 9158 99.40 991 99.36 8167
Battery Level 2.45 191 90.95 191 0 0
Current Time 0.41 31 100 31 0 0
Peripheral Preferred Connection Parameters 99.90 1051 99.90 1037 100 14
Software Revision String 97.04 1017 97.68 1010 50 7
Hardware Revision String 95.95 996 96.58 989 50 7
Serial Number String 97.12 979 97.79 972 50 7
Firmware Revision String 94.99 835 95.72 828 50 7
System ID 97.51 822 98.31 815 50 7
PnP ID 98.02 741 98.02 741 0 0
Peripheral Privacy Flag 99.47 561 100 561 0 0
IEEE 11073-20601 Regulatory Certification Data List 99.64 550 99.64 550 0 0
Reconnection Address 0.56 3 0.56 3 0 0
Central Address Resolution 98.08 307 98.84 85 97.80 222
HTTP Entity Body 100 42 0 0 100 42
Body Sensor Location 70 28 100 28 0 0
Alert Level 7.50 3 8.57 3 0 0
Alert Notification Control Point 100 30 100 30 0 0
Heart Rate Control Point 3.33 1 3.33 1 0 0
Report 14.29 2 14.29 2 0 0
URI 100 14 0 0 100 14
Report Map 42.86 6 42.86 6 0 0
Uncertainty 7.14 1 0 0 7.14 1
HID Information 42.86 6 42.86 6 0 0
Altitude 100 14 0 0 100 14
HTTP Headers 100 14 0 0 100 14
Location Name 100 14 0 0 100 14
RSC Feature 100 9 100 9 0 0
RSC Measurement 11.11 1 11.11 1 0 0
Sensor Location 100 8 100 8 0 0
Protocol Mode 25.00 2 25.00 2 0 0
Tx Power Level 60 3 100 3 0 0
Resolvable Private Address Only 100 4 100 4 0 0
Scan Refresh 33.33 1 33.33 1 0 0
Boot Mouse Input Report 50 1 50 1 0 0
CSC Feature 100 1 100 1 0 0

Fingerprinting Bluetooth-Low-Energy Devices Based on the Generic Attribute Profile

More Related Content

What's hot (20)

Similar to Fingerprinting Bluetooth-Low-Energy Devices Based on the Generic Attribute Profile (20)

More from degarden (20)

Recently uploaded (20)

Fingerprinting Bluetooth-Low-Energy Devices Based on the Generic Attribute Profile