| Internet-Draft | Finding Tracking Tags | September 2024 |
| Fossaceca & Rescorla | Expires 2 April 2025 | [Page] |
Lightweight location tracking tags are in wide use to allow users to locate items. These tags function as a component of a crowdsourced tracking network in which devices belonging to other network users (e.g., phones) report which tags they see and their location, thus allowing the owner of the tag to determine where their tag was most recently seen. This document defines the protocol by which devices report tags they have seen and by which owners look up their location.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://ekr.github.io/draft-fossaceca-dult-finding/draft-fossaceca-dult-finding.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-fossaceca-dult-finding/.¶
Discussion of this document takes place on the Detecting Unwanted Location Trackers Working Group mailing list (mailto:unwanted-trackers@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/unwanted-trackers/. Subscribe at https://www.ietf.org/mailman/listinfo/unwanted-trackers/.¶
Source for this draft and an issue tracker can be found at https://github.com/ekr/draft-fossaceca-dult-finding.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 2 April 2025.¶
Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
DISCLAIMER: This draft is work-in-progress and has not yet seen significant (or really any) security analysis. It should not be used as a basis for building production systems.¶
Lightweight location tracking tags are a mechanism by which users can track their personal items. These tags function as a component of a crowdsourced tracking network in which devices belonging to other network users (e.g., phones) report on the location of tags they have seen. At a high level, location tracking this works as follows:¶
Tags ("Accessories") broadcast an advertisement payload containing accessory-specific information. The payload also indicates whether the accessory is separated from its owner and thus potentially lost.¶
Devices belonging to other users ("Non-Owner Devices" or "Finder Devices") observe those payloads and if the payload is in a separated mode, reports its location to some central service ("Crowdsourced Network").¶
The owner ("Owner Device") queries the central service ("Crowdsourced Network") for the location of their accessory.¶
A naive implementation of this design exposes users to considerable privacy risk. In particular:¶
If accessories simply have a fixed identifier that is reported back to the tracking network, then the central server is able to track any accessory without the user's assistance, which is clearly undesirable.¶
Any attacker who can guess or determine a tag ID can query the central server for its location.¶
An attacker can surreptitiously plant an accessory on a target and thus track them by tracking their "own" accessory.¶
Section 6 provides a more detailed description of the desired security privacy properties, but briefly, we would like to have a system in which:¶
Nobody other than the owner of an accessory would be able to learn anything about the location of a given accessory.¶
It is possible to detect when an accessory is being used to track you.¶
It is not possible for accessories that do not adhere to the protocol to use the crowdsourced network protocol.¶
It is not possible for unverified accessories to use the crowdsourced network protocol.¶
A number of manufacturers have developed their own proprietary tracking protocols, including Apple (see [WhoTracks] and [Heinrich]), Samsung (see [Samsung]), and Tile, CUBE, Chipolo, Pebblebee and TrackR (see [GMCKV21]), with varying security and privacy properties.¶
This document defines a cryptographic reporting and finding protocol which is intended to minimize the above privacy risks. It is intended to work in concert with the requirements defined in [I-D.detecting-unwanted-location-trackers], which facilitate detection of unwanted tracking tags. This protocol design is based on existing academic research surrounding the security and privacy of bluetooth location tracking accessories on the market today, as described in [BlindMy] and [GMCKV21] and closely follows the design of [BlindMy].¶
This work has been inspired by the negative security and privacy implications that were introduced by lightweight location tracking tags, and defined in part by [I-D.detecting-unwanted-location-trackers]. The full threat model is described in detail in [DultDoc4], however, a significant element of the threat model lies in part with the security of the Crowdsourced Network, which will be discussed in detail here.¶
In addition to its designed uses, the Crowdsourced Network also provided stalkers with a means to track others by planting a tracking tag on them and then using the CN to locate the tracker. Thus, this document outlines the requirements and responsibilities of the Crowdsourced Network to verify the authenticity of the participants, while also preserving user privacy.¶
First, the Crowdsourced Network should ensure that only authentic Finding Devices are sending reports to the Crowdsourced Network, and this should occur via an authenticated and encrypted channel. This will help prevent malicious actors from interfering with location reporting services.¶
Second, the Crowdsourced Network should ensure that only authorized Owner Devices are able to download location reports, and this should occur via an authenticated and encrypted channel. This will prevent malicious actors from unauthorized access of location data.¶
Third, the Crowdsourced Network must follow basic security principles, such as storing location reports in an encrypted manner¶
(The benefits of this requirement are self explanatory.)¶
Fourth, the Crowdsourced Network must validate that the accessory registered to an owner is valid. This wil prevent malicious actors from leveraging counterfeit accessories.¶
Fifth, users should should be able to opt-out of their devices participating in the Crowdsourced Network.¶
There is substantial research into stalking via the FindMy protocol and overall crowdsourced network protocol deficiencies have been described in multiple bodies of work, such as:¶
and others.¶
There are some suggested improvements, such as the security properties
described in [GMCKV21] above. The authors of [GMCKV21] also
suggested fusing a private key into the ACC to make it more difficult
to spoof, and requiring that location updates be signed.¶
[Heinrich] and [WhoTracks] pointed out early deficiencies in the
protocol, which [BlindMy] set out to solve. By introducing a Blind
Signature scheme, the authors sought to overcome an attacker
leveraging a large amount of keys to avoid triggering the
anti-tracking framework. In this implementation, keys were
predetermined for a set interval, and signed by the server, such that
a specific, presigned key can only be used during a pre-determined
interval. The drawback of this approach is that the authentication is
left to the OD and the CN, and the CN does not do any
authentication with the FD, so it still could store forged location
reports. Additionally, the FD does not do any authentication with
the ACC, which means that it could potentially interact with
counterfeit ACC devices.¶
[Beck] introduces the idea of Multi-Dealer Secret Sharing (MDSS) as a privacy preserving protocol that should also be considered.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
Section 1.2 of [I-D.detecting-unwanted-location-trackers] provides definitions of the various system components.¶
Accessory (ACC): This is the device which will be tracked. It is assumed to lack direct internet access and GPS, but will possess Bluetooth Low Energy capabilities, which it uses to send advertisement messages. The accessory protocol is defined in [DultDoc3].¶
Advertisement: This is the message that is sent over the BLE Protocol from the Accessory¶
Crowdsourced Network (CN): This is the network that provides the location reporting upload and download services for Owner Devices and Finder Devices.¶
Finder Device (FD): This is a device that is a non-owner device that contributes information about an accessory to the crowdsourced network.¶
Owner Device (OD): This is the device which owns the accessory, and to which it is paired. There can be multiple owner devices, however, the security of that implementation is outside of the scope of this document.¶
Figure 1 provides an overall view of the protocol.¶
In this protocol, the Accessory communicates to Finder Devices or
FDs(such as phones) solely via Bluetooth, and the FDs communicate
to a centralized service on the Crowdsourced Network CN. Only during
the setup phase is the Owner Device OD able to act as a relay
between the Accessory ACC and the Crowdsourced Network CN. In this
implementation, the CN is able to act as a verifier and signer by
leveraging Blind Signatures, which allows the OD to obtain a
signature from the signer CN without revealing the input to the
CN.¶
╭――――――――――――――――――╮
│ o │
│ ╭──────────────╮ │
│ │ │ │
│ │ │ │ .-~~~-.
│ │ │ │ .- ~ ~-( )_ _
o o │ │ │ │ / ~ -.
o o │ │ │ │ | \
o o -------> │ │ │ │ -------> \ .'
o o │ │ │ │ ~- . _____________ . -~
o o │ │ │ │
o o │ │ │ │
│ │ │ │
│ ╰──────────────╯ │
│ (_) │
╰――――――――――――――――――╯
Accessory BLE Finder Device Location CN Server
Advertisement Upload
As part of the setup phase (Section 5.3) the accessory and
owning device are paired, establishing a shared key SK
which is known to both the accessory and the owning device.
The rest of the protocol proceeds as follows.¶
The accessory periodically sends out an advertisement which contains
an ephemeral public key Y_i where i is the epoch the key is valid
for (e.g., a one hour window). Y_i and its corresponding private key
X_i are generated in a deterministic fashion from SK and the epoch
i (conceptually as a X_i = PRF(SK, i)).¶
In order to report an accessory's location at time i a non-owning
device FD encrypts it under Y_i and transmits the pair
( E(Y_i, location), Y_i ) to the central service CN.¶
In order to locate an accessory at time i, the owner uses SK to
compute (X_i, Y_i) and then sends Y_i to the central service.
The central service responds with all the reports it has for Y_i,
and the owner decrypts them with X_i.¶
This design provides substantially improved privacy properties over a naive design:¶
Nobody but the owner can learn the reported location of an
accessory because it is encrypted under Y_i. This includes
the central service, which just sees encrypted reports.¶
It is not possible to correlate the public keys broadcast
across multiple epochs without knowing the shared key SK,
which is only know to the owner. However, an observer who
sees multiple beacons within the same epoch can correlate
them, as they will have the same Y_i. However, fast key
rotation also makes it more difficult to detect unwanted
tracking, which relies on multiple observations of the
same identifier over time.¶
However, there are a number of residual privacy threats, as described below.¶
If the central server is able to learn the identity of the device reporting an accessory or the identity of the owner requesting the location of an accessory, then it can infer information about that accessory's behavior. For instance:¶
If device A reports accessories X and Y owned by different users and they both query for their devices, then the central server may learn that those users were in the same place, or at least their accessories were.¶
If devices A and B both report tag X, then the server learns that A and B were in the same place.¶
If the central server is able to learn where a reporting device is (e.g., by IP address) and then the user queries for that accessory, then the server can infer information about where the user was, or at least where they lost the accessory.¶
These issues can be mitigated by concealing the identity and/or IP address of network elements communicating with the central server using techniques such as Oblivious HTTP [RFC9458] or MASQUE [RFC9298].¶
The detection mechanisms described in [I-D.detecting-unwanted-location-trackers] depend on correct behavior from the tracker. For instance, Section 3.5.1 of [I-D.detecting-unwanted-location-trackers] requires that accessories use a rotation period of 24 hours when in the "separated" state:¶
When in a separated state, the accessory SHALL rotate its address every 24 hours. This duration allows a platform's unwanted tracking algorithms to detect that the same accessory is in proximity for some period of time, when the owner is not in physical proximity.¶
However, if an attacker were to make their own accessory that was
generated the right beacon messages or modify an existing one, they
could cause it to rotate the MAC address and public key Y_i more
frequently, thus evading detection algorithms. The following section
describes a mechanism which is intended to mitigate
this attack.¶
Because evading detection requires rapidly changing keys, evasion can be made more difficult by limiting the rate at which keys can change. This rate limiting works as follows:¶
Instead of allowing the accessory to publish an arbitrary
key Y_i it instead must pre-generate a set of keys,
one for each time window.¶
During the setup/pairing phase, the accessory and owning
device interact with the central service, which
signs each temporal key using a blind signature scheme.
The owning device stores the signatures for each key Y_i.¶
When it wishes to retrieve the location for a given accessory the owning device provides the central service with the corresponding signature, thus proving that it is retrieving location for a pre-registered key; the central service will refuse to provide results for unsigned keys.¶
Note that this mechanism does not prevent the accessory from broadcasting arbitrary keys, but it cannot retrieve location reports corresponding to those keys.¶
This is not a complete defense: it limits an attacker who owns a single accessory to a small number of keys per time window, but an attacker who purchases N devices can then use N times that many keys per window, potentially coordinating usage across spatially separated devices to reduce the per-device cost. [[OPEN ISSUE: Can we do better than this?]]¶
This section provides a detailed description of the DULT Finding Protocol.¶
The there are 5 stages that will be outlined, taking into account elements from both [BlindMy] and [GMCKV21]. These stages are as follows:¶
1) Initial Pairing / Accessory Setup¶
In this phase, the Accessory ACC is paired with the Owner Device
OD, and verified as authentic with the Crowdsourced Network CN¶
2) Accessory in Nearby Owner Mode¶
In this phase, the Accessory ACC is within Bluetooth range of the
Owner Device OD. In this phase, Finder Devices FDs SHALL NOT
generate location reports to send to the Crowdsourced Network
CN. The Accessory SHALL behave as defined in [DultDoc3]. [[OPEN
ISSUE: Need to make sure that walking around with an AirTag in Nearby
Mode does not allow for stalking]]¶
3) Accessory in Separated (Lost) Mode¶
In this phase, the Accessory ACC is not within Bluetooth raange fo
the Owner Device OD, therefore, the accessory must generate "lost"
messages to be received by Finder Devices FD, as described in
[DultDoc3].¶
4) Finder Device creates a location report¶
Finder Device FD receives a Bluetooth packet, and uploads a location
report to the Crowdsourced Network CN if and only if it is confirmed
to be a valid location report.¶
[[OPEN ISSUE: Should this be confirmed by the FD, or the CN? or Both?]]¶
[[OPEN ISSUE: Should there be auth between FD and ACC as well as
FD and CN]]¶
5) Owner Device queries the Crowdsourced Network¶
Owner Device OD queries the Crowdsourced Network CN for the
encrypted location report.¶
[[OPEN ISSUE: Which blind signature scheme to use.]]¶
In order to verify the parties involved in the protocol, we rely on a partially blind signature scheme. [RFC9474] describes a blind signature scheme as follows:¶
The RSA Blind Signature Protocol is a two-party protocol between a client and server where they interact to compute sig = Sign(sk, input_msg), where input_msg = Prepare(msg) is a prepared version of the private message msg provided by the client, and sk is the private signing key provided by the server. See Section 6.2 for details on how sk is generated and used in this protocol. Upon completion of this protocol, the server learns nothing, whereas the client learns sig. In particular, this means the server learns nothing of msg or input_msg and the client learns nothing of sk.¶
The Finding Protocol uses a partially blind signature scheme in which
the signature also covers an additional info value which is not
kept secret from the signing server.¶
During the pairing process, the Accessory ACC pairs with the Owner
Device OD over Bluetooth. In this process, the ACC and OD must
generate cryptographically secure keys that will allow for the OD to
decrypt the ACC location reports.¶
Upon the initial pairing of the the ACC and OD, before the key
generation process, the OD must facilitate communication with the
CN to verify the authenticity of the ACC.¶
The precise details of this communication are implementation-dependent,
but at the end of this process the CN must be able to verify that:¶
For instance, each ACC might be provisioned with a unique serial
number which is digitally signed by the manufacturer, thus allowing
the CN to verify legitimacy. The CN could use a database of
registered serial numbers to prevent multiple registrations.
Once registration is complete, there must be some mechanism for
the OD to maintain continuity of authentication; this too is
implementation specific.¶
The ACC must periodically be provisioned with new temporal
keys which FDs can then use to encrypt reports. Each temporal key
is associated with a given timestamp value,¶
Once the ACC has been authorized, the ACC (or OD on its behalf)
needs to generate its temporal encryption keys Y_i. It then generates
a signing request for the blinded version of each key.¶
contains two values:¶
An opaque string representing the key to be signed, computed as below.¶
The time value for the first time when the key will be used in seconds since the UNIX epoch¶
blindedKey = Blind(pk, Y_i, info)¶
With the following inputs:¶
The public key for CN¶
The temporal key to be signed¶
The timestamp value serialized as an unsigned 64-bit integer in network byte order.¶
Prior to signing the key, the CN must ensure the acceptability of the timestamp.
While the details are implementation dependent, this generally involves
enforcing rate limits on how many keys can be signed with timestamps
within a given window. Once the CN is satisfied with the submission
it constructs a blind signature as shown below and returns it to the OD.¶
[[OPEN ISSUE: Is it safe for ACC to hold all of the precomputed keys? Or does this create a privacy issue? ]]¶
BlindSign(sk, blindedKey, info)
¶
With the following inputs¶
Upon receiving the signed blinded key, the OD unblinds the signature
and stores it. If the OD generated Y_i, it must also transfer it
to the ACC. Note that ACC does not need a copy of the signature.¶
After pairing, when the Accessory ACC is in Bluetooth range of OD,
it will follow the protocol as decribed in [DultDoc3].¶
After pairing, when the Accessory ACC no longer in the Bluetooth
range of OD, it will follow the protocol as decribed below:, which
should correspond to the behavior outlined in [DultDoc3]:¶
ACC periodically sends out an Advertisement which contains the then
current ephemeral public key Y_i. The full payload format of the
Advertisement is defined in [DultDoc3].¶
The Finder Device FD receives the advertisement via Bluetooth. FD
should have a mechanism by which to authenticate that this is a valid
public key with CN. *¶
In order to report an accessory's location at time i, FD extracts
the elliptic curve public key from the advertisement, and records it
own location data, a timestamp, and a confidence value as described in
[Heinrich].¶
FD performs ECDH with the public key Yi and uses it to
encrypt the location data using HPKE Seal [RFC9180]. It sends the
result to the CN along with the hash of the current public key and the
current time. [[OPEN ISSUE: Should we work in terms of hashes or the
public keys. What we send has to be what's looked up.]]. CN stores
the resulting values indexed under the hash of the public key.¶
ODs can retrieve the location of a paired ACC by querying the CN.¶
In order to query for a given time period i it presents:¶
The public key Y_i [or hash of the public key]¶
The CN's signature over Y_i as well as the associated
info value.¶
The CN then proceeds as follows:¶
Verify the signature over the key [hash]¶
Verify that the timestamp in the info value is within an
acceptable period of time (e.g., one week) from the current time
[[OPEN ISSUE: Why do we need this step?]]¶
Retrieve all reports matching the provided Y_i¶
Remove all reports which have timestamps that are not within the acceptable time use window for the key, as indicated by the key's timestamp.¶
Return the remaining reports to OD.¶
Finally, OD uses HPKE Open to decrypt the resulting reports,
thus recovering the location data for report.¶
TODO Security - as described in [DultDoc4]?. This section still mostly needs to be written.¶
The blind signature mechanism described here (adapted from [BlindMy]) helps to limit the damage of noncompliant devices.¶
Because the CN will only generate signatures when the request is
associated with a valid device, an attacker cannot obtain a key
directly for a noncompliant device. However, this does not necessarily
mean that the attacker cannot provision noncompliant
devices. Specifically, if the OD sees the public keys (it need not
know the private keys, as described below) when they are sent to the
CN for signature, then it can provision them to a noncompliant
device.¶
Even an attacker who can provision invalid devices can only obtain one key per time window per valid device. Because key use windows overlap, it is possible to rotate keys more frequently than the window, but in order to rotate keys significantly more frequently than this, the attacker must purchase multiple devices. However, they may be able to provision the keys from multiple valid devices onto the same device, thus achieving a rotation rate increase at linear cost.¶
Note that enforcement of this rate limit happens only on the CN: the
FD does not check. An attacker can generate advertisements with
unsigned keys -- and thus at any rotation rate it chooses -- and the
FD will duly send valid reports encrypted under those keys. The CN
will store them but because the attacker will not be able to produce
valid signatures, they will not be able to retrieve those reports.¶
As noted above, the ACC does not need to prove that it knows the
corresponding private keys for a given public key. The ACC
simply broadcasts the public keys; it is the OD which needs to
know the private keys in order to decrypt the reports.¶
This document has no IANA actions.¶
TODO acknowledge.¶