Home » CTU Improving Post Compromise Security in Practice Questions

CTU Improving Post Compromise Security in Practice Questions

Clone Detection in Secure Messaging:
Improving Post-Compromise Security in Practice
Cas Cremers
Jaiden Fairoze
CISPA Helmholtz Center for Information Security
Saarbrücken, Germany
cremers@cispa.saarland
CISPA Helmholtz Center for Information Security
Saarbrücken, Germany
jfairoze@student.unimelb.edu.au
Benjamin Kiesl
Aurora Naska
CISPA Helmholtz Center for Information Security
Saarbrücken, Germany
benjamin.kiesl@cispa.saarland
CISPA Helmholtz Center for Information Security
Saarbrücken, Germany
s8aunask@stud.uni-saarland.de
ABSTRACT
We investigate whether modern messaging apps achieve the strong
post-compromise security guarantees offered by their underlying
protocols. In particular, we perform a black-box experiment in
which a user becomes the victim of a clone attack; in this attack,
the user’s full state (including identity keys) is compromised by
an attacker who clones their device and then later attempts to
impersonate them, using the app through its user interface.
Our attack should be prevented by protocols that offer postcompromise security, and thus, by all apps that are based on Signal’s
double-ratchet algorithm (for instance, the Signal app, WhatsApp,
and Facebook Secret Conversations). Our experiments reveal that
this is not the case: most deployed messaging apps fall far short of
the security that their underlying mechanisms suggest.
We conjecture that this security gap is a result of many apps
trading security for usability, by tolerating certain forms of desynchronization. We show that the tolerance of desynchronization
necessarily leads to loss of post-compromise security in the strict
sense, but we also show that more security can be retained than
is currently offered in practice. Concretely, we present a modified version of the double-ratchet algorithm that tolerates forms of
desynchronization while still being able to detect cloning activity.
Moreover, we formally analyze our algorithm using the Tamarin
prover to show that it achieves the desired security properties.
CCS CONCEPTS
• Security and privacy → Formal security models; Logic and
verification; Privacy-preserving protocols.
KEYWORDS
security protocols, secure messaging, clone detection, formal verification, double ratchet, post-compromise security, forward secrecy
ACM Reference Format:
Cas Cremers, Jaiden Fairoze, Benjamin Kiesl, and Aurora Naska. 2020. Clone
Detection in Secure Messaging: Improving Post-Compromise Security in
CCS ’20, November 9–13, 2020, Virtual Event, USA
© 2020 Copyright held by the owner/author(s). Publication rights licensed to ACM.
This is the author’s version of the work. It is posted here for your personal use. Not
for redistribution. The definitive Version of Record was published in Proceedings of the
2020 ACM SIGSAC Conference on Computer and Communications Security (CCS ’20),
November 9–13, 2020, Virtual Event, USA, https://doi.org/10.1145/3372297.3423354.
Practice . In Proceedings of the 2020 ACM SIGSAC Conference on Computer and
Communications Security (CCS ’20), November 9–13, 2020, Virtual Event, USA.
ACM, New York, NY, USA, 15 pages. https://doi.org/10.1145/3372297.3423354
1
INTRODUCTION
The advent of modern secure messaging, and the widespread deployment of the Signal protocol library in particular, has brought
modern security mechanisms to millions of users. Specifically, this
includes the double-ratchet algorithm [23] and the security properties it can provide, such as post-compromise security [6], which
provides security guarantees even after a communicating party’s
secrets have been leaked.
As a result, the Signal protocol and many variations have been
studied extensively in the literature [1, 2, 4, 5, 10, 19]. These studies
have shown that (a) the security properties are subtle, and can be
strengthened in various ways, but more importantly, (b) the underlying design of the Signal protocol provably achieves a strong
form of post-compromise security: if Blake’s complete state (including encryption and signing keys) is compromised by an attacker
at some point, but Alex and Blake have a successful “healing” exchange afterwards, the attacker is locked out of the communication
again.
The Signal protocol library can achieve this strong property
through an intricate mechanism called the double ratchet. Despite the additional engineering complexity stemming from the
double ratchet, the Signal protocol library is successfully used in
many modern messaging systems, such as WhatsApp, the Signal
app, Facebook’s Secret Conversations, and Skype. Moreover, other
messengers—such as Viber and Wire—use custom implementations
of the Signal protocol. Together, these apps represent the vast majority of secure-messaging users. This suggests that these users
now enjoy strong post-compromise security guarantees.
For this work, we investigated whether this is indeed the case,
i.e., whether these modern messaging apps do indeed provide the
security guarantees offered by the double ratchet. To do so, we first
performed a black-box study of secure-messaging apps with respect
to one of the weakest possible threats against post-compromise
security, namely clone attacks. In such an attack, Evan compromises
Blake’s state by cloning Blake’s device. At some point in the future,
Evan then tries to attack Blake by running the messaging app from
the cloned state in order to (a) learn messages exchanged by Blake
and their partner Alex, or (b) inject messages into the communication. Now, if an app does indeed provide post-compromise security,
and if Alex and Blake manage to heal before the attack, this attack
should be impossible; in other words, post-compromise security
should ensure the absence of clone attacks after healing.
We establish that surprisingly, in practice, several of the Signalbased apps offer only partial (or even no) post-compromise security
against these relatively simple attacks, in multiple dimensions: some
apps provide no practical guarantees for messages post-healing,
whereas others provide some form of warning to their users. This
suggests that some apps gain very little to no additional security
from the complex mechanisms in the Signal protocol library.
We conjecture that these reduced guarantees are the result of
mechanisms that trade security for usability: the double ratchet requires intricate synchronization, and real-world deployments may
struggle to reliably update and maintain state. If this state is lost,
or synchronization fails for some reason, a strict post-compromise
security mechanism would block the conversation. While this may
be acceptable for security-sensitive users, apps with a wide demographic may instead choose to tolerate desynchronization at the
cost of reduced security. In fact, we know from discussions with
developers of a popular app that this is the case for their app.
In the second half of this work, we explore the theoretical foundations for this apparent dilemma and provide a solution that tolerates
forms of desynchronization while still preventing post-compromise
security threats such as clone attacks.
In particular, we show that a messaging system that tolerates
certain types of non-malicious synchronization failures cannot
achieve full post-compromise security guarantees. We identify and
formalize common non-malicious desynchronization behaviors,
from which we infer how to design mechanisms that tolerate them
without losing all post-compromise security guarantees. We provide
a concrete mechanism that detects clone attacks, and formally prove
its properties using an automated prover. The soundness proof for
our mechanism relies on the assumption that messages arrive in
order, meaning that we trade the double-ratchet’s message-loss
resilience for clone detection.
The main contributions of this paper:
• We conduct an experiment showing that popular messaging
apps—in contrast to their underlying protocols—do not provide post-compromise security even under a simple devicecloning attack. We analyze in detail how this impacts the
security of messages communicated during and after healing.
• We demonstrate that a protocol cannot achieve post-compromise security if it tolerates certain forms of desynchronization of the communicating parties.
• We propose a desynchronization-tolerant version of Signal’s double-ratchet algorithm that allows users to detect
when their partner was cloned. For example, for apps that
prioritize usability over security, this allows a light-weight
response in the specific case of a cloning attack, without
introducing warnings for other common failures; for apps
with security-savvy users, detecting a clone could lead to a
recommendation to generate a new identity key. We analyze
our proposed algorithm formally and prove that it achieves
sound clone detection.
Our formal models are available from https://github.com/dr-clonedetection/model.
The rest of this paper is structured as follows. In Section 2, we
present necessary background as well as related work. In Section 3,
we present our black-box experiment and discuss its outcomes in
detail before we then propose our approach for clone detection—a
modified version of the double ratchet—and its formal analysis in
Section 4. Finally, we summarize our findings in Section 5.
2 BACKGROUND AND RELATED WORK
2.1 Related Security Properties
We briefly review the notions of post-compromise security and forward secrecy, which are achieved by Signal’s double-ratchet algorithm. Intuitively, the former captures security guarantees after a
party was compromised, whereas the latter captures security guarantees before a party was compromised, as illustrated in Figure 1.
Post-Compromise Security. The notion of post-compromise security, also known as backward secrecy, future secrecy, and channel
healing, informally says that a party A has a security guarantee
about communication with another party B, even after B’s secrets
have been compromised [6]. Note that post-compromise security
does not say that A will have a security guarantee immediately after
the compromise; in actual protocols that achieve post-compromise
security, A has a security guarantee once some kind of healing has
occurred after the compromise of B’s secrets.
Even though this might seem unintuitive at first, there are in
fact protocols that achieve forms of post-compromise security. As
with forward secrecy, how post-compromise security is realized
in practice depends on the precise security guarantees a protocol
aims to achieve, and in particular, on the actual secrets that are
compromised. Two typical cases are session-key compromise and
full local-state compromise.
In the case of session-key compromise, only ephemeral cryptographic material is leaked to the attacker. Here, post-compromise
security can be provided by a key-evolving scheme, i.e., a mechanism
that computes the session key using some secret information from
the previous session. For instance, given a key-derivation function
KDF (i.e., a one-way function that derives one or more secret keys
from its input), the 𝑖-th session key sk𝑖 could be computed from
the previous session key sk𝑖−1 and a token 𝑡𝑖 . Here, the token 𝑡𝑖
could, for example, be a shared secret established by the two honest
parties via a Diffie-Hellman key exchange. If the session key is
evolved in this way, the attacker cannot compute a future session
key with only the knowledge of the current session key.
In the case of full local-state compromise, all local user data,
including long-term keys, is leaked to the attacker. Thus, achieving
post-compromise security is much harder in practice. In fact, in
the context of secure messaging, it requires the use of public-key
cryptography for the derivation of message keys [10, 19]. This is
because any key-evolving scheme based on symmetric cryptography satisfies the following: given the previous state and all publicly
exchanged information, it is possible to compute the next state.
Forward Secrecy. The notion of forward secrecy, also known as
perfect forward secrecy, informally says that ciphertexts that were
sent or received prior to the compromise of a party remain secure
compromise
post-compromise security
test
(t)
test
forward secrecy
compromise
(t)
Figure 1: Post-compromise security and forward secrecy.
after the compromise [3, 15]. For instance, imagine A and B are
exchanging encrypted messages until some timepoint 𝑡 at which
an attacker compromises A by stealing A’s secret keys—forward
secrecy guarantees that the attacker will not be able to decrypt any
of the messages exchanged between A and B before timepoint 𝑡.
The exact mechanism used to provide forward secrecy depends on
the level of granularity at which a protocol aims to achieve forward
secrecy. Typical levels of granularity are per message or per session.
In the per-message case, forward secrecy can be achieved by
evolving the encryption keys for every sent message. This is commonly achieved by deriving new keys from old keys via a one-way
function. If keys are regularly updated, messages that were encrypted with previous keys stay secure even if the current key is
compromised. Note that other mechanisms that achieve forward
secrecy have been proposed in the literature, such as puncturable
encryption [8, 14, 16] and time-based methods [9].
In the per-session case, the communicating partners exchange
ephemeral session keys in typical protocols. Using public-key cryptography, a fresh session key is generated independently of the
previous session key, and thus cannot be computed given future
session keys. To reduce the expensive public-key operations, parties
can derive an initial session key and, similar to the per-message
case, evolve it throughout the conversation, e.g., by using a keyderivation function.
2.2
The Double Ratchet
We give a brief overview of Signal’s double-ratchet algorithm for
secure messaging; for full details, we refer to the official documentation [23] or some of the protocol’s proofs [1, 5]. As the name
already indicates, the double ratchet uses two separate ratchets—the
so-called Diffie-Hellman ratchet and the symmetric-key ratchet.
Before the start of a conversation, the two parties, A and B,
initialize a session by establishing a common secret called the root
key. This is achieved with a separate key-agreement protocol that
is not part of the double ratchet; in Signal, the so-called Extended
Triple Diffie-Hellman (X3DH) protocol is used [20]. As part of the
initialization, B also computes a private share 𝑥 and a corresponding
public share 𝑔𝑥 , and then sends the public share to A. The private
and public shares will be important for the Diffie-Hellman ratchet
later on. Once they have finished the initialization, A and B can start
exchanging messages. The idea is that every message is encrypted
and decrypted with a distinct message key, which is derived by
both parties from a so-called symmetric chain key that itself is
derived from the root key. Thus, there is a hierarchy of keys (root
key > symmetric chain key > message key) whereby the DiffieHellman ratchet takes care of updating the root key and computing
initial symmetric chain keys, whereas the symmetric-key ratchet
takes care of updating the symmetric chain keys and deriving new
message keys as illustrated in Figure 2.
Diffie-Hellman Ratchet. At the highest level of the key hierarchy,
a party updates the root key using the Diffie-Hellman ratchet every
time they switch between sending and receiving messages. For example, if A has just received one or more messages and then decides
to send messages, A first updates their root key; after that, A can
then send multiple messages without updating the root key, until
they switch to receiving messages again. Likewise, on the opposite
side, B switches from sending to receiving and thus updates the
root key accordingly.
For the update of the root key, a party first generates a new
private share 𝑦 and a corresponding public share 𝑔 𝑦 . Using their
partner’s current public share 𝑔𝑥 (which they know either from
the initialization or from their partner’s last message), they then
compute a new shared secret (𝑔𝑥 ) 𝑦 = 𝑔𝑥 𝑦 . After this, they input
both the shared secret and the current root key rk𝑖−1 into a keyderivation function KDF, which returns two values, namely the
new root key rk𝑖 and a new initial symmetric chain key sk𝑖 as
⟨rk𝑖 , sk𝑖 ⟩ = KDF (rk𝑖−1, 𝑔𝑥 𝑦 ). Thus, for every update of the root key,
some secret information is added (via the shared secret) to achieve
post-compromise security. Importantly, the initial symmetric chain
key, 𝑠𝑘𝑖 , is then used in the symmetric-key ratchet, which takes us
one level lower in the key hierarchy.
Symmetric-Key Ratchet. The symmetric chain key is used to
derive the actual message keys via the symmetric-key ratchet. In
particular, whenever a party sends or receives a new message, they
derive a distinct message key mk𝑖 from the symmetric chain key and
then also update the symmetric chain key. This is done by simply
applying a KDF again: ⟨sk𝑖+1, mk𝑖+1 ⟩ = KDF (sk𝑖 ). For example,
suppose A has just updated their root key with the Diffie-Hellman
ratchet and thereby derived an initial symmetric chain key 𝑠𝑘 0 . To
encrypt a new message, A now derives a message key mk 1 and
a new symmetric chain key sk 1 from the initial symmetric chain
key by computing ⟨sk 1, mk 1 ⟩ = KDF (sk 0 ). For their next message,
A then derives ⟨sk 2, mk 2 ⟩ = KDF (sk 1 ). A can now repeatedly do
this until they decide to switch from sending to receiving again.
Likewise, on the other side, B will derive the message keys in the
exact same way. By evolving the symmetric chain keys and the
message keys with the KDF, forward secrecy is achieved because
the KDF is a one-way function.
Additional Notes on the Shared Secrets. Note that every message
must contain the current sender’s public share 𝑔 𝑦 to allow the
receiver to compute the shared secret 𝑔𝑥 𝑦 required to update the
root key. The two parties thus take turns incorporating new secret
shares into the root key. To see this, suppose A (who knows B’s
public share 𝑔𝑥 after initialization) starts as a sender. A derives a
new root key rk 1 from the initial root key rk 0 by first generating
a pair ⟨𝑦, 𝑔 𝑦 ⟩, then computing the shared secret 𝑔𝑥 𝑦 , and finally
applying the key-derivation function: ⟨rk 1, 𝑠𝑘 1 ⟩ = KDF (rk 0, 𝑔𝑥 𝑦 ).
When B receives A’s message, B can also compute 𝑟𝑘 1 by computing
the shared secret 𝑔𝑥 𝑦 from 𝑔 𝑦 and B’s own private share 𝑥. If B now
decides to send a message, B generates ⟨𝑧, 𝑔𝑧 ⟩ and then computes
⟨rk 2, 𝑠𝑘 2 ⟩ = KDF (rk 1, 𝑔 𝑦𝑧 ). Consequently, B has replaced the 𝑥part of the initial shared secret 𝑔𝑥 𝑦 by 𝑧. In A’s next turn, A would
then replace the 𝑦-part by a new secret share, and so on—the shared
secret is thus continuously updated with new secret information
from either party.
A
mk 1.1 mk 1.2
mk 3.1 mk 3.2 mk 3.3
message keys:
sk 1.1
symmetric-key ratchet:
A
sk 1.2
sk 3.1
sk 3.2
sk 3.3
𝐸 (𝑚 1 )mk1.1
𝐸 (𝑚 2 )mk1.2
𝐸 (𝑚 3 )mk2.1
rk 2
rk 4
𝐸 (𝑚 4 )mk2.2
Diffie-Hellman ratchet:
rk 1
B
B
symmetric-key ratchet:
𝐸 (𝑚 5 )mk2.3
rk 3
sk 2.1
sk 2.2
sk 2.3
sk 4.1
message keys:
mk 2.1 mk 2.2 mk 2.3
mk 4.1
𝐸 (𝑚 6 )mk3.1
𝐸 (𝑚 7 )mk3.2
𝐸 (𝑚 8 )mk3.3
𝐸 (𝑚 9 )mk4.1
Figure 2: Key structure of the double ratchet during a conversation. Left: Evolution of the keys. Right: A’s view of the conversation (𝐸 (𝑚)mk denotes the encryption of message 𝑚 with message key mk). All keys can be computed by both participants,
allowing the receiver to decrypt messages encrypted by the sender. Whenever participants switch the roles of sender and receiver, they update the root key rk with the Diffie-Hellman ratchet. The symmetric-key ratchet computes a new symmetric
chain key sk and a new message key mk whenever a new message is sent.
Note that the mechanism we have described here, where exactly
one public share is appended to messages, is the one used by the
official double-ratchet algorithm of the Signal protocol. With this
mechanism, if an attacker compromises the full state of a victim, it
learns the victim’s private share, meaning that at least one round
of back-and-forth between the honest communication parties is
required to heal the channel (since the attacker does then not know
the new private share). The Wickr messenger performs more aggressive ratcheting, making the parties exchange more than one public
share per message. This has a number of security implications: the
messages sent by the victim will be immediately private as long
as a fresh pair of shares is sampled at encryption time. However,
the messages received by the victim are not immediately private,
because the attacker knows the pre-shared share pairs exchanged
before compromise. The time required for healing after compromise
then depends on the exact protocol specification—Wickr does not
explicitly specify this parameter in their technical white paper [18]
or on their website (https://wickr.com/).
In theory, the double ratchet enables strong security guarantees:
even if the full state with the identity key is compromised, the
protocol can regain security. In practice, the security of the (identity
key) storage varies widely: from secure storage on modern phones
to substantially less secure storage on older hardware (as used in
poorer regions) or desktop clients.
2.3
Related Work on Clone Detection
After a compromise, if the attacker becomes active and impersonates the victim, there is a chance that their actions create discord
with those of the victim. This discord could be identified and then
used to notify the victim that there was a clone attack. The existing
proven solutions in clone detection use exactly this idea of forcing the attacker to leave traces of their actions, assuming that the
attacker does not merely eavesdrop but injects their own messages.
One way to achieve this, suggested by Yu et al. [29], is with a
third party that records all honest activity. The idea behind the
mechanism is that both parties send certificates of their ephemeral
encryption keys to an append-only log maintainer. Parties then
query the log maintainer to verify the received certificates from
their partner and to obtain proof that no entries of their own have
been added or deleted from the log.
Suppose an attacker compromises a party and uses all existing
ephemeral keys whose certificates are already published in the log.
If they want to continue the impersonation they have to add their
own keys to the log maintainer. This in turn leads to clone detection
once the victim queries the log and detects new entries by another
author. In addition, if the log maintainer is malicious, the parties
can detect the maintainer’s activities by “gossiping” among each
other about the current status of their conversation log.
The drawback of this scheme is that the log maintainer becomes
a bottleneck that requires a form of gossiping unless it is trusted.
Thus, it is desirable to have a protocol that allows for clone detection
without a trusted entity. Such solutions—based on counters, hashing,
or commitments—were proposed by Milner et al. [22].
In their counter-based mechanism, the parties store the number
of messages exchanged with their partner, and append the number
to each message they send. Upon receiving a message, they increment their local value and compare it with the received value. If
the two numbers do not match, a detection event is raised.
In the hash-based approach, the parties—instead of counting the
messages exchanged—keep a hash chain and evolve it with every
message they send, using the previous value of the hash and a
fresh nonce received from their partner. Again, failure to verify the
received hash value with the local one will lead to detection.
Finally, in the commitment-based approach, parties tightly couple their current message with their next one by encrypting current
session data with a key pair used to encrypt the next message. This
encrypted session data serves as a commitment, and after successful
decryption of the next message, it gets verified using the same key.
If data verification fails, a clone is detected since the partner broke
its commitment, thus deviating from normal behavior.
2.4
The Tamarin Prover
We use the Tamarin prover [21] for proving the critical properties of our clone-detection approach in Section 4.2. Tamarin is an
automated-reasoning tool for the analysis of complex security protocols. Tamarin operates on the symbolic level, meaning that bit
strings are abstracted to algebraic terms. As Tamarin is particularly
well-suited for modeling complex state machines with loops and
evolving state, it is a natural choice for modeling our modified
version of Signal’s double ratchet.
To formalize a security protocol in Tamarin, we encode the protocol as a collection of multiset-rewriting rules. Once we have
encoded a protocol by a set of such rules, we can specify desired
properties in a guarded fragment of many-sorted first-order logic
Tamarin tries to prove the property by refuting its negation. In
case Tamarin terminates, it either outputs a proof (if the statement
is true) or a counterexample (if the statement is false). A proof is
provided in the form of a proof tree, whereas a counterexample
is provided in the form of a trace, i.e., a sequence of steps that
corresponds to a possible execution of the protocol. Proofs and
counter-examples can be inspected in Tamarin’s GUI.
3
BLACK-BOX EXPERIMENT
In this section, we describe our black-box experimental analysis of
a range of messaging apps. Our goal is to analyze to what degree
different messaging apps offer post-compromise security in the
context of full-state compromise. In general, a protocol that provides post-compromise security should be resilient to a fully active
attacker that controls the network, after the involved parties have
healed from compromise. Since we are performing black-box analysis, we consider a strictly weaker attacker after the healing: the
attacker executes a clone of the compromised state after the healing.
If a protocol does not offer security with respect to this weaker
attacker, it certainly does not offer post-compromise security.
In a nutshell, we consider a situation where two parties, A and B,
use a messaging app to communicate with each other until a point
where B’s device is cloned (amounting to full-state compromise)
so that there are then two seemingly identical clones Bblake and
Bevan . After that, only one of the two clones, Bblake , continues to
communicate with A, until a point where the other clone, Bevan ,
takes over. Once Bevan has taken over, we want to answer the
following questions:
(A) Given Bblake ’s full local state at the time of cloning, can Bevan
read any of the messages that have been exchanged between
Bblake and A?
(B) How does the messaging app react to the cloning situation?
Does it inform the parties about unusual/suspicious behavior?
Are there any errors? Or does the app just continue as if nothing
happened?
We would expect that a messaging app which provides post-compromise security does not leak any messages exchanged between
Bblake and A to Bevan . Moreover, while not strictly implied by postcompromise security, it would improve security if the app made
the parties aware of the suspicious cloning behavior. As we will
see, none of the messaging apps we considered behave that way.
In the following, we first explain the exact setup of our experiment before we analyze its outcomes.
3.1
Experimental Setup
For our evaluation, we selected messengers that explicitly claim
strong security. We distinguish between
• messengers based on the Signal library (which implements
the Signal protocol),
• messengers based on custom implementations of the Signal
protocol, and
• messengers based on protocols different than Signal.
Table 1 gives an overview of all tested messengers, their protocols,
and their version numbers. We discuss them briefly in the following.
Signal-Based Messengers (Signal, WhatsApp, Facebook, and Skype).
Aside from Open Whisper System’s Signal messenger, we considered three other messengers that directly interface with the Signal
protocol. Namely, WhatsApp bases all communications through its
service on the Signal protocol [28]. Moreover, Facebook Messenger
and Skype have distinct features—called Secret Conversations [11]
and Private Conversations [24], respectively—that support end-toend encryption via the Signal protocol. All messengers in this group
should in theory be able to achieve the same security as Signal.
Messengers with Custom Implementations of the Signal Protocol
(Viber, Wire). Viber [27] and Wire [13] both claim that their respective protocols are in-house implementations of the standard Signal
protocol. Therefore, like the messengers from the above group,
Viber and Wire should be able to match Signal’s security.
Messengers Based on Other Protocols (Wickr, Olvid, Threema, and
Telegram X ). Wickr [18], unlike all other assessed messengers,
utilizes more aggressive ratcheting as discussed in Section 2.2.
Olvid [12] is (according to a technical talk) designed to provide
forward secrecy and post-compromise security through a custom
algorithm, even though the official documentation does not explicitly claim a form of post-compromise security. Finally, Threema [26],
and Telegram X [25] do not attempt to achieve any form of postcompromise security; we included them as a form of control.
Considered Devices. We considered Android devices as well as
desktop clients. Our focus was on the Android devices since most
secure-messaging applications target mobile devices. To emulate
these devices, we used the Genymotion Android emulator due to
its powerful cloning features. All emulated devices were modeled
after Google Nexus 5X devices running Android 8.0 (API 26).
We used desktop clients to assess Wickr and Wire, which provide
full support for the desktop. While these messengers had mobile
versions, we used the desktop clients to circumvent compatibility
issues. We ran the applications on Windows 10 virtual machines
created with VirtualBox. To clone a device, we cloned the entire
virtual machine to duplicate the full local state. Note that for Wickr
and Wire, the desktop clients can exist independently of a mobile
device, as first-class devices.
Protocol
App Version
Signal[23]
libsignal
V4.47.7
Signal Library
WhatsApp[7] FB SC[11]
libsignal
libsignal
V2.19.274
239.1.0.17.119
Skype[24]
libsignal
V8.53.0.104
Custom Signal Implementation
Viber[27]
Wire[13]
Closed Source Proteus
V11.6.3.4
3.10.3138.0
Wickr[18]
wickr-crypto-c
5.38.2
Other Protocols
Olvid[12]
Threema[26]
Closed Source threema-msgapi
V0.7.6
V4.11k
Telegram X[25]
MTProto
V0.22.0.1205-x86
Table 1: Tested messengers with their underlying protocols and version numbers.
pre-test
test
post-test
Bblake
active device
B
Bevan
A
𝑡0
𝑡1
time
𝑡2
𝑡3
inactive device
message exchange
A
Figure 3: Phases of our experiment. In the pre-test phase, A and B exchange messages until 𝑡 1 , where B is cloned, resulting in
Bblake and Bevan . In the subsequent test phase, Bblake communicates with A until 𝑡 2 , where Bevan takes over in the post-test phase.
Table 2 shows which of the messages exchanged between A and Bblake in the test phase can be decrypted by Bevan . Table 3
summarizes how the devices of A and Bevan react to the messages exchanged in the post-test phase. Since we are interested
in the healing behavior of apps with respect to confidentiality and authenticity, we interpret Bblake as the honest party that
potentially heals during the 𝑡𝑒𝑠𝑡 phase, and interpret Bevan as the clone.
3.2
Methodology
In our experiment, we consider two communicating parties, A and B.
We use the notation A → B to denote that A sent a message to B.
Moreover, given a sequence 𝑆 of message exchanges, we write
𝑆 𝑛 to denote that 𝑆 is repeated 𝑛 times. For example, we write
(A → B, B → A) 3 to denote that A and B sent messages back and
forth three times in a row.
Our experiment consists of three phases, shown in Figure 3; we
call them the pre-test phase, the test phase, and the post-test phase.
Pre-Test Phase. In the pre-test phase, we establish (for each messenger) a secure channel between the two parties, A and B. This
involves registering a single device for each party with the service.
After registration, we keep the default settings in all aspects except
for so-called read receipts and typing indicators; the former indicate
to the other party that a message has been read, the latter indicate
that the other party is currently typing a message. We disabled
these two features whenever possible, to prevent additional protocol messages from firing. Moreover, we made sure to grant the
apps full device permissions to permit proper app function.
The timepoint 𝑡 0 marks the point where device configuration
is completed. After that, A and B exchange a series of messages
to move the conversation to an “active” state. Specifically, these
pre-test messages are of the form (A → B, B → A) 5 .
Test Phase. The test phase begins with the cloning event at 𝑡 1 .
Right before the cloning event, B is blocked from the network to
ensure that states of the subsequent clones do not automatically
update. We clone B’s device in such a way that one clone, Bblake ,
joins the network while the other clone, Bevan , is kept inactive. This
effectively mimics the continuation of the single device B, since
from the outside it should appear as if just a single device left and
rejoined the network. Once the cloning has been performed, A and
Bblake exchange a few messages. These messages constitute the test
messages, and are of the form (A → Bblake , Bblake → A) 5 .
After the test messages have been exchanged, at timepoint 𝑡 2 , the
two clones switch status—Bevan becomes active and Bblake leaves
the network. At this point, post-compromise security could potentially be violated if the initially-offline device were able to automatically retrieve messages from the test phase.
Post-Test Phase. In the post-test phase, A and Bevan conduct a
final round of message exchanges. We use this phase to investigate
whether or not conversation can continue after Bblake left. For the
actual message exchange, we consider four different sequences of
messages:
(1) (A → Bevan, Bevan → A) 5
(2) (Bevan → A, A → Bevan ) 5
(3) (A → Bevan, A → Bevan, Bevan → A, Bevan → A) 5
(4) (Bevan → A, Bevan → A, A → Bevan, A → Bevan ) 5
For each case, we carry out the corresponding message-exchange
pattern on the devices. Specifically, when Bblake ’s device is brought
online, it is given a period of 30 seconds to allow for asynchronous
update with the messaging platform. During this period, the device
might receive messages from the test phase through some platformspecific mechanism whose details we are not concerned with in
our black-box experiment. Note that some platforms might deliver
only some (or even none) of the messages from the test phase to
Bblake . After this period, we capture the application-level behavior
of both parties.
Before we discuss the results of our experiment, we want to
highlight that the methodology of our experiment is guided by the
double-ratchet’s design, meaning that we set up conditions under
which the double ratchet should provide post-compromise security and then test if this is indeed the case. The reader might want
Message Number
1 2 3 4 5
A → Bblake ✓ ✓ ✓ ✓ ✓
Signal
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✓ ✓ ✓ ✓ ✓
WhatsApp
Signal
Bblake → A ✓ ✓ ✓ ✓ ✓
Library
A → Bblake ✗ ✗ ✗ ✗ ✗
Facebook SC
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✗ ✓ ✓ ✓ ✓
Skype PC
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✓ ✓ ✓ ✓ ✓
Viber
Custom
Bblake → A ✓ ✓ ✓ ✓ ✓
Signal
A → Bblake ✗ ✗ ✓ ✓ ✓
Wire
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✗ ✗ ✗ ✗ ✗
Wickr
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✓ ✓ ✓ ✓ ✓
Olvid
Other
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✓ ✓ ✓ ✓ ✓
Protocols
Threema
Bblake → A ✓ ✓ ✓ ✓ ✓
A → Bblake ✗ ✗ ✗ ✗ ✗
Telegram X
Bblake → A ✗ ✗ ✗ ✗ ✗
Table 2: Messages from the test phase. During the test phase,
A and Bblake exchange messages. Afterwards, the attacker
Bevan activates the clone. For each message from the test
phase, ✓ denotes that the attacker Bevan does not obtain it,
and ✗ denotes that the attacker learns it.
to keep this in mind when interpreting the results of the experiment, especially regarding Olvid and Wickr—two messengers that
are designed to provide post-compromise security under possibly
different conditions.
3.3
test phase. Wire does not protect the second message from A to
Bblake , which contradicts its theoretical level of security.
(3) One-Sided Protection. Facebook Secret Conversations and Wickr
offer what we call one-sided protection: A’s messages to Bblake
are leaked to Bevan , but Bblake ’s messages to A remain secure.
In Wickr’s case, a batch of public-key secrets is known to either
party at any given time, as mentioned in Section 2.2. With this
information, it is feasible that Bevan had all public-key shares
from A and thus could decrypt one side of the test phase messages. Interestingly, while Facebook’s Secret Conversations interfaces with the Signal library directly [11], it does not achieve
post-compromise security.
(4) No Protection. Telegram X is the only messenger where all messages are leaked to the attacker. We expected this, because
Telegram’s MTProto algorithm makes no attempt to provide
post-compromise security. However, as we have seen, a number of platforms without advanced mechanisms for achieving
post-compromise security can still protect against an attacker
restricted to UI-access. Uncovering the reasons for this security
gap may positively influence messaging-protocol design.
Post-Compromise Behavior. The second question (B) of our experiment concerns the behavior of the messengers in the post-test
phase. It turns out that the various messengers react in different
ways to the cloning event and the respective message exchange
from the test phase. A summary of their behavior is given in Table 3.
We remark that the post-compromise behavior of messaging apps
is unrelated to the post-compromise-security property—this section
explores whether the messaging channel is usable after a cloning
attack. We can distinguish three different kinds of behavior, which
are the same for all four types of message exchanges:
Results
As already mentioned, we are interested in the answers to the
following two questions:
(A) Can Bevan read any of the messages that have been exchanged
between Bblake and A in the test phase?
(B) How does a messaging app react to the cloning situation?
For question (A), the outcome of our experiment is summarized
in Table 2. The results can be grouped into four categories:
(1) Full Protection. For some messengers, namely Signal, WhatsApp,
Viber, Olvid, and Threema, all messages exchanged in the test
phase remain secure. In contrast to the other messengers in that
group, Threema does not use a double-ratchet-based algorithm,
nor does it provide a mechanism (such as an asynchronous
ratchet) to enforce post-compromise security—this shows that
messengers can protect from coarse-grained adversaries with
limited technical proficiency (i.e., high-level access to the platform) without complex asynchronous ratcheting.
(2) Near-Full Protection. Skype PC and Wire offer near-full protection—all messages except the first one (Skype PC) or the first
two (Wire) from A to Bblake stay protected. Both messengers use
double-ratchet-based cryptography, which should theoretically
protect all messages except for the very first message from the
(1) No Protection and no Detection: Permit Communication Without
Errors. WhatsApp, Facebook Secret Conversations, Viber, Telegram X, Threema, and Wickr allow A and Bevan to send and
receive messages in the post-test phase, without exhibiting any
errors or showing any error messages. From the participants’
point of view (through the user-interface) it is undetectable that
any form of unusual behavior has occurred.
(2) Partial Protection or Detection. When using Signal, Olvid, or
Wire, A and Bevan seem able to send and receive messages
in the post-test phase, but there are errors about which the
messengers alert them explicitly.
In particular, in Signal, when a sender (either A or Bevan )
sends a message to the other party, then the receiver’s app will
just display the message “Bad encrypted message”. Thus, while
the message is blocked, the sender is not informed.
Wire behaves similarly to Signal: sent messages lead to
corresponding receipt notifications on the sender side, but the
receiver cannot decrypt the messages and instead is shown the
error message “A message from [A|B] was not received” together
with an error code and a link that says “Reset session”.
Finally, Olvid is different from the others in that its protection is one-sided. Messages sent from Bevan are not received by
Post-Compromise
Error msg?
Blocked?
Send Recv
A → Bevan


Signal
Bevan → A


A → Bevan

WhatsApp
Signal
Bevan → A

Library
A → Bevan

Facebook SC
Bevan → A

A → Bevan



Skype PC
Bevan → A



A → Bevan

Viber
Custom
Bevan → A

Signal
A → Bevan


Wire
Bevan → A


A → Bevan

Wickr
Bevan → A

A → Bevan

Olvid
Other
Bevan → A


Protocols
A → Bevan

Threema
Bevan → A

A → Bevan

Telegram X
Bevan → A

Table 3: Post-compromise behavior based on messages from the
post-test phase. After a test phase during which A and Bblake
heal, subsequent communication between A and the attacker
Bevan should become impossible.
The “blocked?” column indicates whether communication is
indeed blocked as expected. If the communication is blocked,
the next two columns indicate whether the sender and recipient obtain information in this case. Here, ✓ indicates that the
sender respectively receiver see an error message.
A, and Bevan ’s device does not display corresponding receipt notifications; however, messages sent from A are still received by
Bevan , and A’s device shows corresponding receipt notifications.
(3) Full Protection and Detection: Halt Communication with Errors.
Skype Private Conversations is unique in that it terminates communication between A and Bevan , and shows an error message
on both devices: right after A attempts to send the first message
to Bevan , Skype will show A the message “Private conversation
ended. Send a new invitation to resume your private conversation with B”, also Bevan will see the analogous message “Private
conversation ended. Send a new invitation to resume your private
conversation with A”.
3.4
Analysis
Our experiment shows that messaging apps behave differently upon
compromise of one of the parties. Two aspects are particularly
relevant in our work—the secrecy of messages sent and received
after the compromise, and the ability of the apps to detect the clone
activity and to notify the parties that suspicious activity took place.
We would expect messengers using the Signal protocol library
(Signal, WhatsApp, Facebook SC and Skype PC) to behave in the
same way with respect to the secrecy of messages in the test phase.
However, as it turns out, they span into three different groups.
Signal and WhatsApp protect all messages in the test phase, as
expected from the double ratchet’s post-compromise security.
Skype PC provides the protection of all messages except for the
first message sent from A to Bblake . A possible reason for this is that
no healing has happened at the point when Bevan takes over (i.e.,
the root key has not been updated via the Diffie-Hellman ratchet);
therefore, the first message A sends to Bblake (and which is later
received by Bevan ) is exactly what Bevan expects.
The most surprising behavior of Signal-based apps is that of
Facebook SC, where the clone, Bevan , can decrypt all messages sent
from A to Bblake . For this to happen, either Bevan would have to
compute the same ratcheting steps as Bblake to generate the exact
same sequence of shared secrets (which happens with negligible
probability), or A would have to re-encrypt the messages for Bevan ,
by using a previous state or by implicitly restarting the conversation.
We conjecture that this is due to an implicit restart of the session,
after which old messages are re-encrypted and sent again.
Differences can also be observed when it comes to the apps’
reactions to the clone behavior. The most drastic reaction is that of
Skype PC, which completely halted the communication in the posttest phase. In contrast, Facebook SC, WhatsApp, and Viber—which
are also Signal-based—continue as if nothing happened. Continuing the conversation in the post-test phase without any errors
would require a mechanism that resynchronizes the states of the
communicating partners. This indicates that having the parties
desynchronized is deemed a normal situation by these apps, which
would explain why no suspicious activity is flagged. Clearly, these
apps are sacrificing security for usability.
Skype’s approach is the more secure solution, though it might
lose on the usability end. However, it would be desirable if parties
could be notified of the reason why the conversation is not private
anymore, thus explicitly detecting possible clones.
Signal ranges somewhere between Skype PC and the other Signalbased apps: Even though the conversation seemingly continues
from the perspective of the sender, none of the messages can be
decrypted. The parties are therefore implicitly forced to restart the
conversation without an explicit explanation what went wrong.
As we show next, security and usability can be reconciled in a
better way, by equipping the double ratchet with a sound clonedetection mechanism that makes no false or vague claims.
4
PROPOSED IMPROVEMENT
As we have seen, existing messaging apps do not achieve the security guarantees offered in theory by their underlying protocols. In
particular, they do not achieve post-compromise security in practice.
Many of these apps rely on the Signal protocol, which has at its core
the well-known double ratchet algorithm, which combines DiffieHellman secret sharing with key chains based on key-derivation
functions (which in practice are hash functions).
In theory, the double ratchet evolves the key chains of both the
sender and the receiver after every exchanged message, to compute
for each message a distinct message key that is used for encryption
on the sender side and for decryption on the receiver side. Thus,
for every sender key chain of one party, there is a corresponding
receiver key chain of the other party. For example, when A sends
messages 𝑚 1, . . . , 𝑚 5 to B, A starts out with a message key 𝐾1 for
encrypting 𝑚 1 and then successively evolves their sender key chain
to compute the message keys 𝐾2, . . . , 𝐾5 for the subsequent messages 𝑚 2, . . . , 𝑚 5 . On the receiver side, B evolves the receiver key
chain accordingly to decrypt the messages sent by A.
In the context of our cloning experiment, the following should
therefore happen in theory: In the test phase, B’s first clone, Bblake ,
evolves the key chains according to the messages exchanged with
A, whereas Bevan (who is offline during the test phase) does not
evolve the key chains at all. Once the test phase is over, the key
chains of Bevan and A are then desynchronized, meaning that an
error should occur once Bevan and A try to exchange messages.
Since our experiment shows that most apps do not behave that
way, it appears that there is a mismatch between the double ratchet’s
specification and its actual implementations in messaging apps. We
conjecture that this mismatch is the result of an accepted trade-off
between security and error tolerance. Unfortunately, this increased
error tolerance provides an attacker with increased capabilities,
opening the door for clone attacks that, as we have seen, violate
post-compromise security. Whereas in theory, an app should be
able to detect the suspicious behavior caused by cloning a device,
this is often not the case in practice.
Possible Causes of Desynchronization Errors. In reality, errors—
which manifest themselves in the form of unexpected message keys
used for encryption and decryption—can have very natural reasons.
Usually, they are the result of desynchronization errors based on
so-called state loss, which causes a device to incorrectly update its
internal state, i.e., the part of memory relevant to the correct functioning of an app. The internal state contains important information
such as secret keys used in the key chains. State loss can therefore
lead to the desynchronization of the key chains maintained by the
sender and the receiver, resulting in unexpected behavior. From
discussions with developers, we find that for global user bases, two
kinds of state loss occur in practice:
• Total-state loss is the event where a party completely erases
any previous secret values. A typical cause of total-state loss
is the loss of a device, leading its owner to reinstall the messaging app on a new device. Similar situations occur when a
device is wiped and reinstalled.
• Single-state loss is the event where a party sends a message and then fails to update its local state due to hardware
malfunction, e.g., because the state update is not properly
propagated to persistent memory. When the party then tries
to send a new message, it uses its previous state to derive
the (old) key and thus continues its key chain from there.
This event is much less likely than the first, but can occur
for older soft- and hardware used in impoverished regions.
Tolerating State Loss Violates Post-Compromise Security. The tolerance of state loss can increase the usability of a messaging app by
allowing users to continue their communication, but this increased
usability can violate post-compromise security. A simple example is
an app that tolerates total-state loss: Suppose A and B are communicating with each other, and then B’s device is cloned by the attacker;
assume that A and B heal afterwards. This should imply the clone is
locked out again. However, the attacker could then pretend to have
suffered from total-state loss and thus reinitialize its state with A to
communicate with A. If then at some point the original B tries to
send a message to A, the message cannot be decrypted by A, which
could in turn lead A to reinitialize with the original B. This means
that the attacker was able to impersonate B by injecting their own
messages and eavesdropping on A’s messages, and they could in
fact even do this repeatedly without being detected, even though A
and B healed. This violates post-compromise security.
Combining Error Tolerance with Clone Detection. As we cannot
achieve post-compromise security if we want to tolerate certain
forms of state loss, a trade-off between error tolerance and security
is inevitable. We therefore aim for a protocol that tolerates state
loss while still being able to perform sound clone detection. Here,
by sound we mean that the protocol only indicates cloning behavior
if there was indeed cloning behavior; in other words, there should
be no false positives. This allows a messaging app to respond to the
cloning activity in a proper way, informing the users and possibly
taking further measures to achieve secure communication again.
In the following section, we present a modified version of Signal’s
double ratchet that achieves exactly that, but before we present our
protocol, we first want to highlight a few important subtleties that
are involved with clone detection in the presence of state loss.
Why State Loss Breaks Existing Approaches for Clone Detection.
There exist various clone-detection solutions (see Section 2.3). However, as none of these mechanisms consider the occurrence of state
loss, they can falsely identify honest actions as actions of a clone.
For instance, in the log maintainer approach discussed in Section 2.3, a party that appends an entry to the log but fails to record
it locally would detect a clone upon requesting a proof of extension
from the log. The same holds for the other mechanisms, where
failure to register one message exchange leads to detection. The
following would happen in each protocol upon single-state loss:
(1) counter-based: the same counter would be reused when sending the next message,
(2) hash-based: the hash value would not be evolved correctly,
(3) commitment-based: the party would fail to remember the
commitment it made.
Clone detection can get even more involved when, for instance,
both parties repeatedly suffer from single-state loss. We thus require
an approach that can deal with these problems.
4.1
Concrete Proposal
We propose a slight modification of Signal’s double ratchet that is
able to detect clones in the presence of both single-state loss and
total-state loss: We add to every message a message counter and
a message authentication code (MAC) computed via a special key,
which we call the epoch key. In the original double ratchet (see
Section 2), whenever two parties start a communication session,
they first agree on a root key (in Signal, via the X3DH protocol),
which is then updated subsequently to evolve the key chains for
sending and receiving messages. We define the epoch key ek to be
a hash of this root key, i.e., given a root key rk and a hash function
ℎ, the epoch key is ℎ(rk).
When two parties start a communication session, they both
compute the epoch key and additionally initialize two counters—the
send counter and the receive counter—to the value 0. Now, whenever
a party sends a message, it appends its send counter as well as a
MAC of the message computed with the epoch key. Upon receipt
of a message with send counter 𝑛, the receiver does two things:
(1) it verifies the MAC with the epoch key, and
(2) it checks if the send counter 𝑛 of the message is greater than
or equal to its own receive counter; if so, it sets its receive
counter to 𝑛.
If the MAC verification or the counter check fails, the receiver
concludes that its supposed communication partner must have been
cloned. Note the receiver does not indicate whether the current
message was sent by a clone of its supposed partner or by the
partner itself; all it knows is that the partner must have been cloned.
Figure 4 shows an example trace of our protocol.
As we will show, the combination of message counters and the
MAC suffices not only to guarantee the soundness of our approach
(lack of false positives) but also allows us to detect a large class of
cloning activities. On the other hand, completeness of detection
(lack of false negatives) is not achievable even in the simpler scenario with no state loss [22]. Due to space constraints, we illustrate
the limitations of detection with concrete examples in Appendix A.
The use of an incrementing counter implies that the soundness
of our approach depends on the loss-less, in-order delivery of messages, which could, for instance, be guaranteed by adding a buffer.
It should be mentioned though that (as noted by Alwen et al. [1])
the addition of a buffer can introduce complications.
Ideas Behind the Approach. The intuitive ideas behind using the
message counters and the MAC are as follows: By adding a message
counter to each message, we allow the receiver to track the last
counter it received from its partner—it does so by setting its receive
counter to 𝑛 for every received counter in Step 2 above. Thus, if
a message with a lower message counter arrives, the receiver can
conclude that something must have gone wrong, because if there
were only one party sending messages, this party would not send
an old message counter. Notice that we deliberately allow message
counters that are equal to the current receive counter to cover the
case where the other party has suffered from single-state loss and
thus sent the same message counter twice in a row.
The message counters alone, however, are not sufficient for guaranteeing soundness, which is why we need the MAC computed
with the epoch key: Suppose A has exchanged a few messages
with Bblake until a point where A’s receive counter and Bblake ’s
send counter both have the value 𝑛. At that point, Bblake ’s device is
cloned and the clone, Bevan , pretends to have suffered from totalstate loss. Then, as discussed on Page 4, A might reinitialize with
Bevan , meaning that they reset their send and receive counters to 0
and compute a new epoch key from the new root key.
The clone Bevan can then exchange messages with A. Now, as
long as Bblake doesn’t send any messages to A, the clone, Bevan ,
will be able to communicate with A without A noticing anything
unusual. However, if Bblake then sends a message to A, the cloning
behavior will be detected by A, because the MAC of Bblake ’s message
cannot be verified as Bblake is still using an old epoch key. Had A
just checked if the message counter sent by Bblake (which still has
B
A
[𝑐𝑖−1 , 𝑀𝑖−1 ]
• Compute the message key mk𝑖 from the
symmetric chain key sk𝑖−1 :
– ⟨sk𝑖 , mk𝑖 ⟩ := KDF (sk𝑖−1 )
• Increment the send counter sc𝐴,𝑖 :
– sc𝐴,𝑖 := sc𝐴,𝑖−1 + 1
• Encrypt the payload 𝑚𝑖 and append the send
counter sc𝐴,𝑖 :
– 𝑐𝑖 := ⟨. . . , {𝑚𝑖 }mk𝑖 , sc𝐴,𝑖 ⟩
• Compute a MAC with the epoch key ek:
– 𝑀𝑖 := MAC ek (𝑐𝑖 )
[𝑐𝑖 , 𝑀𝑖 ]
• Verify the MAC 𝑀𝑖 with the epoch key ek:
– If 𝑀𝑖 ≠ MAC ek (𝑐𝑖 ): DETECT CLONE
• Compare the message counter sc𝐴,𝑖 with the local
receive counter rc𝐵,𝑖 :
– If sc𝐴,𝑖 < rc𝐵,𝑖 : DETECT CLONE • Update the receive counter – rc𝐵,𝑖 := sc𝐴,𝑖 • Derive message key mk𝑖 from the current state: – ⟨sk𝑖 , mk𝑖 ⟩ := KDF (sk𝑖−1 ) • If decryption fails, tolerate single-state loss: – Derive mk𝑖 from the previous state: • ⟨sk𝑖 , mk𝑖 ⟩ := KDF (𝑠𝑘𝑖−2 ) – If decryption fails again, ask partner to resend: • RequestRetry( [𝑐𝑖 , 𝑀𝑖 ]) Figure 4: Example trace of our proposed protocol. Differences to the original double-ratchet algorithm are highlighted in blue. A appends to their original message their current send counter sc𝐴,𝑖 and a MAC 𝑀𝑒𝑘 (𝑐𝑖 ), computed with their epoch key ek, of the ciphertext. When B receives the message, B detects a clone if (a) the MAC does not verify with B’s epoch key ek, or (b) the counter is lower than B’s current receive counter rc𝐵,𝑖 . Otherwise, B updates their receive counter and (a) decrypts the message with their current state (normal message), or, if this fails, (b) decrypts the message with their previous state (single-state loss), or, if this fails too, (c) asks A to send the message again (potential attack or honest message, see Example 4.1). the value 𝑛 + 1) was lower than A’s current receive counter, A might have concluded that 𝑛 + 1 was greater than their current receive counter and thus A wouldn’t have detected a clone. Verification of the MAC with the epoch key thus serves as an additional measure to deal with total-state loss. Our approach adds minimal overhead to the original double ratchet: Each message is extended by just two additional fields—the message counter and the MAC. For example, a message-counter field of 32 bits (allowing to count up to more than 4 billion messages) and a 256 bit MAC add an overhead of 288 bits, or 36 bytes, amounting to three short English words encoded in UTF-16. The space requirements for storing the two counters and the epoch key on devices are also negligible. Our approach is thus much less costly than, e.g., storing all previous message keys to compare newly arrived messages against them. Such an approach would have the additional drawback that it allows an attacker to decrypt all old messages once it obtained access to the keys stored on a device. Additional Considerations. One could reasonably question the necessity of adding the message counters and the MAC instead of just detecting clones based on whether received messages can be encrypted with the current message key of the receiver chain. After all, an honest (non-cloned) party should anyhow encrypt messages with the right message key, and if it does not, there must have been a clone, right? The following example shows that this seemingly intuitive approach would lead to the unsound detection of clones. In the example, keep in mind that whenever a party sends a message right after it received a message, it evolves its root-key chain and uses the new root key to derive a message key for the new message: Example 4.1. Suppose A evolves their root-key chain to send a new message to B, but then suffers from single-state loss and thus forgets that the root-key chain has evolved. When B then receives A’s message, B uses A’s new public key (which is part of their message) to also evolve the root-key chain and decrypt the encrypted portion of A’s message. Usually, A and B would now both have the same root key, but since A forgot theirs, the root keys of A and B don’t match. Now B decides to respond to A by sending a new message. B thus evolves the root-key chain again to derive a new message key to encrypt the message. When A then receives this message, the message cannot be decrypted, but if A concluded that B’s device must have been cloned, A would be wrong. In the above example, B’s message will contain a valid MAC (computed with the epoch key) and a message counter that is greater than A’s current receive counter. Our approach would thus behave as expected and not detect any cloning behavior. Example traces of how parties synchronize after single-state loss and total-state loss, as well as clone detection, can be found in Appendix A. We next present a formal analysis of our proposed algorithm. 4.2 Formal Analysis The goal of our analysis is to create a faithful formal model of our modified double ratchet and to prove the soundness of our proposed clone-detection mechanism using the Tamarin prover. Our model consists of 21 rewriting rules as well as 20 different lemmas (stating, for instance, crucial invariants) that we proved in order to obtain our main statement, which says that whenever an honest user detects that its partner has been cloned, the partner has indeed been cloned. As it is impossible to discuss the entire model within the restricted space of this submission, we focus on a general overview as well as on critical modeling decisions. We uploaded our formal model (including all lemmas and their corresponding proofs as well as documentation to reproduce our results) to the GitHub repository https://github.com/dr-clone-detection/model. Overview of the Model. We allow an unbounded number of users that can create unique bidirectional communication channels with other users. Once a user has established a communication channel with a particular partner, that channel is uniquely identified by a so-called user thread. This means that a single user can start an unbounded number of user threads in general, but only one user thread per communication partner. Once a pair of users has established such a channel, they can initialize (and later reinitialize) a session with each other. The initialization allows them to establish secret keys—in particular the root key of the double ratchet—and to start exchanging messages with each other. As we focus on the double ratchet and not on the initialization protocol (which, in Signal’s case is X3DH) we just modeled an abstract initialization consisting of three rewriting rules. The only assumption underlying the initialization procedure is that a pair of users cannot perform two or more initializations with each other in parallel. As part of the initialization, the users derive the epoch key for the clone-detection mechanism (by applying a hash function to the root key) and initialize their send counters to 1 and their receive counters to 0. After the initialization, one of the users (the initiator) is ready to start sending messages while its partner (the responder) can receive these messages. As defined by the double ratchet, they can then take turns, switching between sending and receiving messages. Whenever a user switches from sending to receiving, it performs a ratcheting step of the Diffie-Hellman ratchet to compute a new public key and in particular a new root key. Moreover, when a user sends or receives multiple messages in a row, it evolves its sender or receiver key chain accordingly by performing ratcheting steps of the symmetric-key ratchet. All this is modeled by dedicated rewriting rules that take care of sending and receiving messages and accordingly updating a user’s state. Clone-Detection Mechanism. We incorporated our clone-detection mechanism by forcing a sending user to update its send counter and to append this send counter along with a MAC computed with the epoch key to every message. The receiver takes care of updating its receive counter with every received message. To detect clones, we added dedicated receiver rules that can be executed when either the MAC of a message is invalid or the receive counter is lower than a user’s current receive counter. State Loss. To capture single-state loss and total-state loss, we added two additional rewriting rules: In the single-state loss case, the rule defines that a user sends a message (like it would do in a normal sender rule), but instead of updating its local state (by evolving its key chains and updating its send counter), it doesn’t alter the state at all. In the total-state loss case, a user can (at any point in time, and not just when sending a message) forget its whole state. From that point on, the user suffering from state loss can do nothing else than reinitialize with its partner. Attacker Capabilities. The attacker in our model can at any point in time take the current state of a user and clone the state. We modeled this with a dedicated rule that takes all the relevant state (e.g., the root key, symmetric chain key, etc.) of a given user thread and spawns a new user thread, which can then act like a normal user and send messages to the communication partner associated with the user thread. This means that the attacker’s capabilities, after cloning a user thread, are restricted to the capabilities of a normal user. This models the assumption that the attacker really just cloned a device and then used a messaging app through its user interface. We modeled this by replacing the usual In and Out facts—offered by Tamarin to model an attacker who can control the network by intercepting or altering general messages on the network—with dedicated Send and Receive rules. We thus also used typed messages in the Receive facts, because the attacker is anyhow not able to create arbitrary messages. Finally, note that once a user has been cloned, its partner is unable to distinguish between messages received from the original user and messages received from the attacker. As discussed earlier, the only restriction is that the messages sent by one user thread must arrive in the order they were sent. This applies only to user threads and not to users: It is possible that a user thread of user Bblake first sends a message, and then the attacker thread of Bblake ’s clone Bevan sends a message, but Bevan ’s message arrives first. Methodology. Our main statement intuitively says that whenever an honest party (i.e., a party that itself is not a clone) claims that its communication partner was a clone, then the partner must indeed have been cloned at an earlier point in time. In guarded first-order logic, the statement reads as follows: ∀ userThread user partner epochKey rcvCtr key t1 . (DetectClone(userThread, user, partner, rcvCtr, messageMAC)@t1 ∧ ¬∃t2 . t2 < t1 ∧ CloneUser(user, partner)@t2 ) ⇒ ∃ t3 . t3 < t1 ∧ CloneUser(partner, user)@t3 As already mentioned, formally proving this seemingly simple statement with Tamarin required us to prove 19 additional lemmas. The main idea behind our proof strategy was to distinguish between two cases that lead a user to detect that its partner was cloned: (1) the partner sends a message with an invalid MAC, or (2) the partner sends a message with a message counter that is lower than the user’s current receive counter. Apart from lemmas that state several invariants and fine-grained properties useful for both cases, we then proceeded as follows. In the case of an invalid MAC, we proved that honest users always know the current epoch key of a communication session and therefore never send an invalid MAC. Thus, if our supposed partner sends an invalid MAC, the partner must have been cloned. In the case of the message counters, we proved (by induction) that within a single communication session, honest users always send messages with monotonically increasing message counters. Thus, if a user receives a message counter that is too low, the monotonicity property is violated, meaning that the partner must have been cloned. Note here that the message counters might not increase strictly monotonically due to single-state loss. Note also that message counters do not increase monotonically across sessions, as the initialization of new sessions resets the message counters . Summary of the Analysis. All 20 lemmas of our formal model can be proved automatically by Tamarin in around 26 minutes on an 8-core machine with 30GB of memory; this, however, required extensive fine-tuning and tweaking of the heuristics for several lemmas. The lemma that took by far the most time (around 19 minutes) is the above-mentioned lemma stating that message counters increase monotonically. The reason why it takes Tamarin so long to prove this lemma is because it has to consider all possible combinations of rewriting rules that refer to message counters, which leads to a quadratic blow-up in the number of rules. Our analysis guarantees that under the assumptions outlined in the beginning of this section, our approach is sound, meaning that it does not yield false positives. As our approach can deal with cloning events from earlier sessions (due to the MAC computed with the epoch key) as well as with cloning events from the current session (due to the message counters), it detects a large range of cloning events. Note that a completeness statement like, if a partner was cloned, the mechanism will detect it cannot hold, as it cannot be guaranteed in general that a clone is distinguishable from the original user. We discuss this in more detail in Appendix A. Finally, as our protocol is a simple extension of Signal’s double ratchet, we believe that our formal model can serve as an excellent foundation for other work that aims at formally analyzing the double ratchet with the Tamarin prover. 5 CONCLUSIONS We have shown that many popular messaging apps do not provide the security guarantees offered by their underlying protocols. In particular, we demonstrated that in practice, cloning a device can lead to surprising behavior that violates post-compromise security. A likely reason for this unexpected reduced security is that the app developers made a trade-off between security and usability, tolerating some forms of desynchronization based on state loss. As we have demonstrated, such a trade-off is inevitable, as full postcompromise security has to be given up if state loss is tolerated. We identified two typical forms of state loss: single-state loss and total-state loss. The former usually happens when a device sends a message but then fails to update its state accordingly (e.g., failed write/flush) and thus forgets that it sent the message. The latter happens, for instance, when a user loses their device and thus reinstalls an app on a new device, which leads to a loss of all formerly established secrets. To reconcile security and usability, we then presented a securemessaging protocol—a modified version of Signal’s double ratchet that offers a sound and simple approach for clone detection. Our approach, which is based on message counters and message authentication codes, enables users to detect cases where their supposed communication partner has been cloned, thus allowing them to react accordingly—for instance, by re-establishing secret keys with their partner via a secure channel, or for security-savvy users, generating a new identity key. To prove the soundness of our approach, we constructed a formal model of our protocol using the automated-reasoning tool Tamarin. The main result of our analysis states that whenever our clonedetection mechanism indicates to a party that its partner has been cloned, then that partner has indeed be cloned. This formal analysis should increase the trust in our protocol, thus making it a more secure approach than simple ad-hoc mechanisms. As discussed in work on usable security [17], users might have trouble interpreting error messages. Further study is thus required to find usable ways of properly incorporating our clone-detection approach into apps’ user interfaces. REFERENCES [1] Joël Alwen, Sandro Coretti, and Yevgeniy Dodis. 2019. The Double Ratchet: Security Notions, Proofs, and Modularization for the Signal Protocol. In EUROCRYPT (1) (Lecture Notes in Computer Science), Vol. 11476. Springer, 129–158. [2] Mihir Bellare, Asha Camper Singh, Joseph Jaeger, Maya Nyayapati, and Igors Stepanovs. 2017. Ratcheted encryption and key exchange: The security of messaging. In Annual International Cryptology Conference. Springer, 619–650. [3] Colin Boyd, Anish Mathuria, and Douglas Stebila. 2020. Protocols for Authentication and Key Establishment, Second Edition. Springer. https://doi.org/10.1007/9783-662-58146-9 [4] Andrea Caforio, F. Betül Durak, and Serge Vaudenay. 2019. On-Demand Ratcheting with Security Awareness. IACR Cryptology ePrint Archive 2019 (2019), 965. [5] Katriel Cohn-Gordon, Cas J. F. Cremers, Benjamin Dowling, Luke Garratt, and Douglas Stebila. 2017. A Formal Security Analysis of the Signal Messaging Protocol. In EuroS&P. IEEE, 451–466. [6] Katriel Cohn-Gordon, Cas J. F. Cremers, and Luke Garratt. 2016. On Postcompromise Security. In IEEE 29th Computer Security Foundations Symposium, CSF 2016, Lisbon, Portugal, June 27 - July 1, 2016. IEEE Computer Society, 164–178. https://doi.org/10.1109/CSF.2016.19 [7] Josh Constine. 2018. WhatsApp hits 1.5 billion monthly users. $19b? not so bad. https://techcrunch.com/2018/01/31/whatsapp-hits-1-5-billion-monthlyusers-19b-not-so-bad. accessed: 2020-05-02. [8] David Derler, Tibor Jager, Daniel Slamanig, and Christoph Striecks. 2018. Bloom filter encryption and applications to efficient forward-secret 0-RTT key exchange. In Annual International Conference on the Theory and Applications of Cryptographic Techniques. Springer, 425–455. [9] Mohammad Sadeq Dousti and Rasool Jalili. 2015. Forsakes: A forward-secure authenticated key exchange protocol based on symmetric key-evolving schemes. Advances in Mathematics of Communications 9, 4 (2015), 471–514. [10] F Betül Durak and Serge Vaudenay. 2019. Bidirectional asynchronous ratcheted key agreement with linear complexity. In International Workshop on Security. Springer, 343–362. Messenger Secret Conversations, Technical Whitepa[11] Facebook. 2017. per. https://fbnewsroomus.files.wordpress.com/2016/07/messenger-secretconversations-technical-whitepaper.pdf. Accessed: 2020-05-02. [12] Matthieu Finiasz and Thomas Baignères. 2019. Security Model of Mobile Messaging Apps. https://olvid.io/assets/slides/2019-06-12%20Olvid%20-%20GDR% 20S%C3%A9curit%C3%A9%20Informatique.pdf. accessed: 2019-12-03. [13] Wire Swiss GmbH. 2018. Wire Security Whitepaper. https://wire-docs.wire.com/ download/Wire+Security+Whitepaper.pdf. accessed: 2020-05-02. [14] Matthew D Green and Ian Miers. 2015. Forward secure asynchronous messaging from puncturable encryption. In 2015 IEEE Symposium on Security and Privacy. IEEE, 305–320. [15] Christoph G. Günther. 1989. An Identity-Based Key-Exchange Protocol. In Advances in Cryptology - EUROCRYPT ’89, Workshop on the Theory and Application of of Cryptographic Techniques, Houthalen, Belgium, April 10-13, 1989, Proceedings (Lecture Notes in Computer Science), Vol. 434. Springer, 29–37. https: //doi.org/10.1007/3-540-46885-4_5 [16] Felix Günther, Britta Hale, Tibor Jager, and Sebastian Lauer. 2017. 0-RTT key exchange with full forward secrecy. In Annual International Conference on the Theory and Applications of Cryptographic Techniques. Springer, 519–548. [17] Amir Herzberg and Hemi Leibowitz. 2016. Can Johnny finally encrypt?: evaluating E2E-encryption in popular IM applications. In Proceedings of the 6th Workshop on Socio-Technical Aspects in Security and Trust, STAST 2016, Los Angeles, CA, USA, December 5, 2016, Gabriele Lenzini, Giampaolo Bella, Zinaida Benenson, and Carrie E. Gates (Eds.). ACM, 17–28. https://doi.org/10.1145/3046055.3046059 [18] Chris Howell, Tom Leavy, and Joël Alwen. 2017. Wickr’s Messaging Protocol. https://wickr.com/wickrs-messaging-protocol/. accessed: 2019-08-16. [19] Joseph Jaeger and Igors Stepanovs. 2018. Optimal channel security against fine-grained state compromise: the safety of messaging. In Annual International Cryptology Conference. Springer, 33–62. [20] Moxie Marlinspike and Trevor Perrin. 2016. The X3DH Key Agreement Protocol. https://signal.org/docs/specifications/x3dh/. accessed: 2019-08-12. [21] Simon Meier, Benedikt Schmidt, Cas Cremers, and David A. Basin. 2013. The TAMARIN Prover for the Symbolic Analysis of Security Protocols. In CAV (Lecture Notes in Computer Science), Vol. 8044. Springer, 696–701. [22] Kevin Milner, Cas Cremers, Jiangshan Yu, and Mark Ryan. 2017. Automatically detecting the misuse of secrets: Foundations, design principles, and applications. In 2017 IEEE 30th Computer Security Foundations Symposium (CSF). IEEE, 203–216. [23] Trevor Perrin and Moxie Marlinspike. 2016. The Double Ratchet Algorithm. https://signal.org/docs/specifications/doubleratchet/. accessed: 2019-08-12. [24] Skype. 2018. Skype Private Conversation, Technical white paper. https://az705183.vo.msecnd.net/onlinesupportmedia/onlinesupport/media/ skype/documents/skype-private-conversation-white-paper.pdf. Accessed: 2020-05-02. [25] Telegram. 2019. End-to-End Encryption, Secret Chats. https://core.telegram.org/ api/end-to-end. Accessed: 2020-05-02. [26] Threema. 2019. Threema. Cryptography Whitepaper. https://threema.ch/pressfiles/cryptography_whitepaper.pdf. Accessed: 2020-05-02. [27] Viber. 2019. Viber Encryption Overview. https://www.viber.com/app/uploads/ viber-encryption-overview.pdf. accessed: 2020-05-02. [28] WhatsApp. 2017. WhatsApp Encryption Overview. Technical white paper. https: //www.whatsapp.com/security/WhatsApp-Security-Whitepaper.pdf. Accessed: 2020-05-02. [29] Jiangshan Yu, Mark Ryan, and Cas Cremers. 2017. Decim: Detecting endpoint compromise in messaging. IEEE Transactions on Information Forensics and Security 13, 1 (2017), 106–118. A EXAMPLE TRACES B A In this appendix, we provide example traces to illustrate how our proposed version of Signal’s double-ratchet algorithm functions during resynchronization procedures and clone detection. In Figure 5, we show an example trace in which a clone is detected. In Figure 6, we show an example trace for recovering from single-state loss. Similarly, we show in Figure 7 an example trace for recovering from a total-state loss. • compute key mk 4 and next chain key 𝑐𝑘𝑠 4 – (𝑐𝑘𝑠 4 , 𝑚𝑘 4 ) := KDF(𝑐𝑘𝑠 3 ) • increment counter : 𝑠𝑐 𝐴,4 := 𝑠𝑐 𝐴,3 + 1 • encrypt 𝑚  4 and append counter 𝑠𝑐 𝐴,4 – 𝑐 4 := ..., {| 𝑚 4 |}𝑚𝑘4 , 𝑠𝑐 𝐴,4 • Compute MAC with epoch key 𝑒𝑘 – 𝑀4 := MAC𝑒𝑘 (𝑐 4 ) [𝑐 4, 𝑀4 ] Bblake A [𝑐𝑖 , 𝑀𝑖 ] • MAC verifies: 𝑀4 = MAC𝑒𝑘 (𝑐 4 ) Compromise (Single-State Loss) • A fails to store the new values after sending the message 𝑚 4 • Counter verifies: 𝑠𝑐 𝐴,4 > 𝑟𝑐 𝐵,3
• derive key 𝑚𝑘 4 from current state
– (𝑐𝑘𝑟 4 , 𝑚𝑘 4 ) := KDF(𝑐𝑘𝑟 3 )
• decryption successful, update receiver counter
– success: 𝑟𝑐 𝐵,4 := 𝑟𝑐 𝐵,3 + 1
[𝑐𝑖+1, 𝑀𝑖+1 ]
[𝑐𝑖+𝑛 , 𝑀𝑖+𝑛 ]
Bevan
• Compute (again) key 𝑚𝑘 4 and next chain
key 𝑐𝑘𝑠 4
– (𝑐𝑘𝑠 4 , 𝑚𝑘 4 ) := KDF(𝑐𝑘𝑠 3 )
• increment counter : 𝑠𝑐 𝐴,4 := 𝑠𝑐 𝐴,3 + 1
• compute key mk𝑖+1 from compromised state
– (𝑐𝑘𝑠𝑖+1, 𝑚𝑘𝑖+1 ) := KDF(𝑐𝑘𝑠𝑖 )
• increment counter : 𝑠𝑐 𝐵,𝑖+1 := 𝑠𝑐 𝐵,𝑖 + 1
• encrypt 𝑚𝑛𝑒𝑤,𝑖+1
 and append counter 𝑠𝑐 𝐵,𝑖+1

– 𝑐𝑛𝑒𝑤,𝑖+1 := …, {| 𝑚𝑛𝑒𝑤,𝑖+1 |}𝑚𝑘𝑖+1 , 𝑠𝑐 𝐴,4
• Compute MAC with epoch key 𝑒𝑘
– 𝑀𝑛𝑒𝑤,𝑖+1 := MAC𝑒𝑘 (𝑐 4 )
• encrypt new message 𝑚𝑛𝑒𝑤,4 and append
counter 𝑠𝑐 𝐴,4

– 𝑐𝑛𝑒𝑤,4 :=
…, {| 𝑚𝑛𝑒𝑤,4 |}𝑚𝑘4 , 𝑠𝑐 𝐴,4
• Compute MAC with epoch key 𝑒𝑘
– 𝑀𝑛𝑒𝑤,4 := MAC𝑒𝑘 (𝑐𝑛𝑒𝑤,4 )
[𝑐𝑛𝑒𝑤,4, 𝑀𝑛𝑒𝑤,4 ]
[𝑐 1, 𝑀1 ]
• MAC verifies: 𝑀𝑛𝑒𝑤,4 = MAC𝑒𝑘 (𝑐𝑛𝑒𝑤,4 )
• Counter verifies: 𝑠𝑐 𝐴,4 = 𝑟𝑐 𝐵,4
• MAC verifies: 𝑀𝑛𝑒𝑤,4 = MAC𝑒𝑘 (𝑐𝑛𝑒𝑤,4 )
• Verify Counter: DETECT CLONE
– 𝑠𝑐 𝐵,i+1 < 𝑟𝑐 𝐴,i+n • derive key 𝑚𝑘 5 from current state – (𝑐𝑘𝑟 5 , 𝑚𝑘 5 ) := KDF(𝑐𝑘𝑟 4 ) • decryption fails, handle Single-State Loss • derive 𝑚𝑘 5 from previous old state – (𝑐𝑘𝑟 4 , 𝑚𝑘 5 ) := KDF(𝑐𝑘𝑟 3 ) • decryption successful, don’t increment counter – 𝑟𝑐 𝐵,𝑖 := 𝑟𝑐 𝐵,𝑖−1 Figure 5: Example trace of clone detection. Bblake ’s state gets compromised after sending message 𝑚𝑖 . Bblake continues the conversation with A exhanging 𝑛 more message until the attacker Bevan becomes active and injects a message of its own. A detects a clone, since the counter received 𝑠𝑐 𝐵,𝑖+1 is smaller than the local counter 𝑟𝑐 𝐴,𝑖+𝑛 Figure 6: Example trace of the parties synchronizing after A had single-state loss. A sends message 4, B receives the message and updates their state. A has single-state loss after sending the message, thus it will perform the same computations to send the next message. B receives again, detects singlestate loss and computes the message key from the previous state, synchronizing again with A. B A [𝑐𝑖 , 𝑀𝑖 ] ( Total-State Loss ) • A loses all the state values Key Agreement Protocol (A has the shared secret 𝑟𝑘 0 and B’s ephemeral public key 𝐸𝑘𝐵 and own key pair (𝑒𝑘𝐴 , 𝐸𝐾𝐴 )) • store the new epoch key 𝑒𝑘 of the session – 𝑒𝑘 := HASH (𝑟𝑘 0 ) • initialize sender and receiver counter – 𝑠𝑐 𝐴,0 := 0 ; 𝑟𝑐 𝐴,0 := 0 • compute the new root key and a sending chain key from initial secret and DH output 𝑒𝑘 – (𝑟𝑘 1 , 𝑐𝑘𝑠 0 ) := KDF(𝑟𝑘 0 , 𝐸𝐾𝐵 𝐴 ) • compute key 𝑚𝑘 1 from sending chain 𝑐𝑘𝑠 0 • encrypt message 𝑚 1 and append own public key 𝐸𝑘𝐴, incremented counter 𝑠𝑐 𝐴,1  – 𝑐 1 := ..., {| 𝑚 1 |}𝑚𝑘1 , 𝐸𝑘𝐴 , 𝑠𝑐 𝐴,1 • Compute MAC with epoch key 𝑒𝑘 [𝑐 1, 𝑀1 ] (B has the shared secret 𝑟𝑘 0 and own key pair (𝑒𝑘𝐵 , 𝐸𝐾𝐵 )) • store the new epoch key 𝑒𝑘 of the session – 𝑒𝑘 := HASH (𝑟𝑘 0 ) • initialize sender and receiver counter – 𝑠𝑐 𝐵,0 := 0 ; 𝑟𝑐 𝐵,0 = 0 Previous work [22] has shown that detection completeness—i.e., providing a clone-detection approach that can detect all kinds of cloning behavior—is not achievable. A trivial example to illustrate this is an attacker that, after cloning a device’s state, does not use the device. Another example is an attacker whose actions do not interfere with those of the victim, as in a stateless protocol. In both cases, clone detection based solely on the messages exchanged in a protocol is provably impossible. As we have seen, allowing state loss gives the attacker the additional capabilities of (1) resetting the state of a conversation between its victim and a third party and (2) using the same key twice to encrypt a message. This activity can be detected as described in detail in 4.1. However, even if the attacker simply follows the protocol, there are two major limitations: First, consider a cloned party that suffers from total state loss and has yet to restart the conversation with its partner. Until that happens, the attacker can impersonate the victim using the cloned state and go undetected. This comes as no surprise, considering that state is not carried between the two sessions before and after state loss from the honest party. Thus, the protocol acts as if it was stateless and, by previous results, detection is not possible. A possible trace is the following: (1) Bevan : Compromised (2) Bevan → A, A → Bevan (3) Bevan : Total-State Loss restart (4) Bblake −−−−−−−→ A, A → Bblake restart (5) Bevan −−−−−−−→ A (No Detection) , A → Bevan , • verify MAC and counter value of received message • get 𝐸𝑘𝐴 from message, compute a new root key and receiving chain key from initial secret and DH output 𝑒𝑘 – (𝑟𝑘 1 , 𝑐𝑘𝑟 0 ) := KDF(𝑟𝑘 0 , 𝐸𝐾𝐴 𝐵 ) • derive message key, decrypt the message and increment counter – (𝑐𝑘𝑟 1 , 𝑚𝑘 1 ) := KDF(𝑐𝑘𝑟 0 ) – (𝑟𝑐 𝐵,1 := (𝑟𝑐𝐵, 0 + 1) Second, an attacker that clones a device and then immediately (before the state of the conversation is updated) sends messages while claiming single-state loss does not get detected. The reason is that to the receiver, it looks like its peer is suffering from repeated single-state loss. To recover, the honest parties need to perform an asymmetric-ratchet step, thus achieving post-compromise security and locking the attacker out. key k Figure 7: Example trace of the parties synchronizing after total-state loss. A sends message 𝑖, then has total-state loss (e.g., A loses their phone). When A wants to communicate again, first it will establish a new common secret with B, derive the new root key, initialize the counters and then compute the message key. B upon receiving the message will follow the same procedure and then the parties are once again synchronized. B UNDETECTABLE ATTACK TRACES In this section, we discuss scenarios in which clone detection is hard or even impossible. (1) Bevan −−−−−→ A (2) Bevan : Compromised & Single-State Loss key k key k (3) Bblake −−−−−→ A, Bblake −−−−−→ A, ... (4) A → Bevan new key k’ (5) Bevan −−−−−−−−−−→ A (No Detection) , A → Bevan , To conclude, when the cloned party loses state, detection is hard or even impossible until the key is forwarded or a new one is derived. However, this class of possible attacks is not very realistic in our current setting. In order for the attacker to go undetected, it would have to either guess exactly when state loss happens to its victim or control the victim’s hardware and induce the state-loss event itself. Both cases are hard to achieve for a simple cloning attacker. Assignment Support Read the paper “Clone Detection in Secure Messaging: Improving Post-Compromise Security in Practice” (see attached): 1. Summarize the paper and describe the main problem that the paper is trying to solve. 2. Explain in your own words what “Post-Compromise Security” means in this paper. 3. Describe the threat model for these attacks. Hint: It might not be explicitly stated but think about what capabilities and goals the presumed attacker has. 4. Explain what the results of the black-box experiment show and what we can learn from them. 5. Describe the proposed improvement (message counters). Explain what difference this change makes and how (to what extent) it solves the problem

Place your order
(550 words)

Approximate price: $22

Calculate the price of your order

550 words
We'll send you the first draft for approval by September 11, 2018 at 10:52 AM
Total price:
$26
The price is based on these factors:
Academic level
Number of pages
Urgency
Basic features
  • Free title page and bibliography
  • Unlimited revisions
  • Plagiarism-free guarantee
  • Money-back guarantee
  • 24/7 support
On-demand options
  • Writer’s samples
  • Part-by-part delivery
  • Overnight delivery
  • Copies of used sources
  • Expert Proofreading
Paper format
  • 275 words per page
  • 12 pt Arial/Times New Roman
  • Double line spacing
  • Any citation style (APA, MLA, Chicago/Turabian, Harvard)

Our guarantees

Delivering a high-quality product at a reasonable price is not enough anymore.
That’s why we have developed 5 beneficial guarantees that will make your experience with our service enjoyable, easy, and safe.

Money-back guarantee

You have to be 100% sure of the quality of your product to give a money-back guarantee. This describes us perfectly. Make sure that this guarantee is totally transparent.

Read more

Zero-plagiarism guarantee

Each paper is composed from scratch, according to your instructions. It is then checked by our plagiarism-detection software. There is no gap where plagiarism could squeeze in.

Read more

Free-revision policy

Thanks to our free revisions, there is no way for you to be unsatisfied. We will work on your paper until you are completely happy with the result.

Read more

Privacy policy

Your email is safe, as we store it according to international data protection rules. Your bank details are secure, as we use only reliable payment systems.

Read more

Fair-cooperation guarantee

By sending us your money, you buy the service we provide. Check out our terms and conditions if you prefer business talks to be laid out in official language.

Read more

Order your essay today and save 30% with the discount code ESSAYHELP