Deconstructing the Decentralization Trilemma
Harry Halpin
a
Inria de Paris, 2 Rue Simone Iff, Paris, France
Keywords:
Software Architecture, Security, Privacy, Scalability, Federation, Decentralization, Blockchain.
Abstract:
The vast majority of applications at this moment rely on centralized servers to relay messages between clients,
where these servers are considered trusted third-parties. With the rise of blockchain technologies over the
last few years, there has been a move away from both centralized servers and traditional federated models
to more decentralized peer-to-peer alternatives. However, there appears to be a trilemma between security,
scalability, and decentralization in blockchain-based systems. Deconstructing this trilemma using well-known
threat models, we define a typology of centralized, federated, and decentralized architectures. Each of the
different architectures has this trilemma play out differently. Facing a possible decentralized future, we outline
seven hard problems facing decentralization and theorize that the differences between centralized, federated,
and decentralized architectures depend on differing social interpretations of trust.
1 INTRODUCTION
Although there has been a move towards decentral-
ization, projects with decentralized architectures have
had fundamental difficulties: Bitcoin and Ethereum
seem to be unable to scale to as large a number of
transactions as centralized systems such as Visa. On
the other hand, centralized projects are increasingly
the subject of data leakage and other attacks, calling
their security into question. This has been phrased
as the “decentralization trilemma” by the co-founder
of Ethereum, Vitalik Buterin, and is a widely spread
truism in blockchain development that has not been
rigorously analyzed and critiqued.
1
Intuitively, there does seem to be fundamental
trade-offs between decentralization, scalability, and
security as illustrated in Figure 1: Systems can be
less decentralized and more scalable versus more de-
centralized and less scalable. Indeed, decentraliza-
tion does seem like a trade-off against scalability, but
most large real-world deployments of scalable soft-
ware, such as Amazon, are actually distributed sys-
tems with large trust assumptions and centralized co-
ordination (DeCandia et al., 2007). Also, there are se-
cure centralized systems that use advanced encryption
(such as the Signal instant messenger), and decen-
tralized systems that have been found to be insecure,
as various attacks on the distributed hash tables used
a
https://orcid.org/0000-0003-2143-6965
1
https://github.com/ethereum/wiki/wiki/Sharding-FAQ
by peer-to-peer file-sharing networks show (Wolchok
et al., 2010). Decentralization can be thought of as
arising from a separate threat model than traditional
security assumptions: A lack of trust in centralized
servers. In this paper, we outline the threat model,
the malicious server, that these decentralized archi-
tectures are trying to address in Section 2. Decen-
tralization, which we define in terms of an adversar-
ial approach to distributed systems (Troncoso et al.,
2017), is then explored in Section 3 as separate from
classical security and scalability requirements. Rather
than a binary division, we view decentralization on a
spectrum that can be broadly construed as centralized,
federated, and decentralized architectures in Section
4, and we provide an analysis in Section 5. In Section
6 we outline six open problems that decentralized sys-
tems face in meeting the requirements currently met
by centralized architectures. In our conclusion in Sec-
tion 7, we reassess decentralization and turn to the so-
cial hypothesis at the heart of decentralization.
2 THE MALICIOUS SERVER
In distributed systems, all entities are considered to be
capable of sending messages (Lamport et al., 1982).
In centralized systems, users do not directly receive
messages but are mediated by a client (a device, a pro-
gram such as a mail-reader or a browser, etc.) where
the client communicates to a server that stores and for-
Halpin, H.
Deconstructing the Decentralization Trilemma.
DOI: 10.5220/0009892405050512
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - SECRYPT, pages 505-512
ISBN: 978-989-758-446-6
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
505
Figure 1: The decentralization trilemma.
wards messages to the client. As exemplified by cloud
computing, this server is assumed to be always online.
One of the primary advantages of centralized servers
seems to be that the deployment and upgrading of any
protocol is easier via the usage of a centralized server.
As the server mediates all messages, the server
is a trusted third party. Centralized systems can
be secure and maintain privacy against even power-
ful adversaries. Secure messaging applications rang-
ing from Signal to WhatsApp, depend on centralized
servers (Unger et al., 2015). In the case of secure
messaging protocols, even if the message content is
encrypted, the server is usually necessary for delivery,
especially if the client is offline. As shown by the sim-
ple case that messages are assumed not be dropped by
the server, the security and privacy properties of cen-
tralized servers are dependent on a single root of trust.
Under the adversary model of malicious security,
trusted components of a larger distributed process
may no longer follow the protocol and so no longer
maintain their security properties (Yung, 2015). To
achieve malicious security, the protocol must work
even with a component no longer maintains its se-
curity properties. Although originally aimed at cryp-
tographic primitives, malicious security also applies
at the level of the entire architecture of a system. In
particular, a malicious server is one that no longer se-
curely relays messages according to the protocol.
We theorize that the goal of decentralization is
to achieve protocols with malicious security in the
face of any component, and so also achieve resis-
tance in terms of the byzantine failure of components,
where components behavior is unrestricted, and so
may be malicious or “unintentionally” faulty (such
as being offline) (Lamport et al., 1982). Given the
popular deployment of servers in computing proto-
cols, amongst all possible components, the primary
threat model of decentralization is a malicious server,
where the server is untrusted. Malicious security can
be achieved in a decentralized protocol simply by not
relying on a server at all. Malicious servers are a
real-world adversary: As the revelations by Edward
Snowden showed that it was trivial to compromise
the security of e-mail if they were not end-to-end en-
crypted by simply retrieving the cleartext from the
centralized server. Even if the message content is en-
crypted, “metadata” such as the time and recipients
of the message, often has no legal protection from
surveillance (Slobogin, 2014).
The goal of a malicious server in the centralized
client-server setting is to read messages and possi-
bly forge the message content of one or more clients.
Even if the message is encrypted, the malicious server
may be “honest but curious” and so also has the goal
of determining the identity of the senders and recip-
ients of messages, i.e. the discovery of the “social
graph” of the communication. The server may use
techniques like dropping messages, replaying mes-
sages, or sending fake messages to identify users. The
server is local, it does not have the ability to observe
the entire network, but can observe every message
it relays, and so differs from threat models like the
global passive adversary that can observe all mes-
sages in an entire network or an adversary that can
arbitrarily corrupt any component in the system. The
malicious server may be a mobile adversary and so
the corruption of the server may be limited in dura-
tion (Yung, 2015).
Before cloud computing, servers were generally
viewed as trustworthy and message-passing between
them governed primarily by standardized protocols
such as SMTP and HTTP, leading to federated sys-
tems. Although the servers were generally regarded
as trusted rather than malicious in early federated in-
ternet architectures, issues such as spam and a need
for payments led to increased centralization. The
pendulum is now swinging in the other direction: It
is our fundamental hypothesis that decentralized sys-
tems arose in response to the threat model of the mali-
cious server. For example, Bittorrent came into usage
after the the centralized index of Napster was elimi-
nated, and Bitcoin became popular after the central-
ized servers in classical e-cash schemes were viewed
as a liability after the “take down” of e-Gold (Tron-
coso et al., 2017).
3 PROPERTIES: SECURITY AND
SCALABILITY
Security and scalability are very broad notions. It is
useful to decompose them into more distinct and bet-
ter known technical properties that can be enforced
via cryptography and engineering. The list of proper-
ties below is somewhat arbitrary, but reflects classical
distinctions within the fields of computer security and
SECRYPT 2020 - 17th International Conference on Security and Cryptography
506
distributed systems.
3.1 Security Properties
A number of traditional of security properties are
needed to counter a malicious server. Security prop-
erties should be able to hold in terms of messaging
even in the presence of a malicious server, which is
considered the “adversary” below:
Confidentiality: The adversary cannot read the
cleartext of any message.
Integrity: The adversary cannot alter the message
without detection.
Authenticity: Only the intended recipient can re-
ceive and send the message (in order to prevent
replay and impersonation attacks by the server).
Although informally thought of as “security”
properties in the decentralization trilemma, we can
consider privacy properties separately:
Unlinkability: Each message of a client is unlink-
able from any given other message by the client
or any other client (Pfitzmann and Hansen, 2005).
Unobservability: A message cannot be distin-
guished from not sending a message (Pfitzmann
and Hansen, 2005).
3.2 Scalability Properties
Additional properties relate to the scalability of the
system, and so can be contrasted with security and
privacy requirements (Das et al., 2018). A system
should be able to scale so it is both available to clients
when needed, can support new clients, and deliver
messages within a period of time acceptable to the
user.
Availability: Messages can be sent or received by
the client when requested.
Capacity: New clients may be added at any time
while maintaining availability.
Latency: Messages may be sent by one client and
received by another client within an acceptable
period of time.
Some requirements cannot be fulfilled by tradi-
tional cryptographic and engineering requirements,
as these related to the freedom of users, particularly
when they are under attack by a malicious server.
Transparency: Messages or information about
message-passing behavior are recorded for clients
to access.
Figure 2: From left to right: (1) Centralized (2) Federated
(3) Decentralized.
Portability: Messages may be send and received
across all servers and clients, so a user may move
their messages to another server or send the mes-
sage directly themselves in case their server be-
comes malicious.
These properties are by no means exhaustive,
but are meant to deconstruct and ground the terms
“security” and “scalability” in the decentralization
trilemma in well-accepted technical properties. There
are more informal properties that can fit within this
framework. Take usability, which can be defined as
the ability to easily retrieve and send messages by the
user.
4 DECENTRALIZED
ARCHITECTURES
In order to minimize the damage a malicious server
could cause, decentralization is put forward as an al-
ternative architecture to trusted third parties. How-
ever, there are multiple variations of architectures
claiming to be decentralized: For example, efforts
like Autocrypt claim that e-mail, despite involving a
SMTP server, is decentralized.
2
So defining and com-
paring these architectures in terms of messaging pass-
ing between clients and servers is necessary. We di-
vide the three primary architectural choices as central-
ized, federated, and decentralized. These architecture
were classically illustrated by Baran in the following
Figure 2 (Baran et al., 1964).
3
2
https://autocrypt.org/
3
In his original work, Baran called these (1) Centralized,
(2) Decentralized, and (3) Distributed, but terminology has
been changed to be consistent with current software archi-
tecture practice where “federated” has a clear meaning in
terms of a multiple authoritative servers with one or more
clients each, while the difference between “decentralized”
and “distributed” in informal modern parlance has become
vague.
Deconstructing the Decentralization Trilemma
507
4.1 Centralized
An architecture is strictly centralized if there is a
single trusted third-party that mediates all messages.
Clients are restricted to a server. For example using
a messaging service such as WhatsApp, for any two
users, all messages must be passed through the What-
sApp server. Slack, Facebook, Signal, Apple Pay-
ments, and other popular cloud-based services are ex-
amples of centralized architectures.
4.2 Federated
In the federated architecture, client may pass mes-
sages to different servers, and these servers may com-
municate to pass messages between clients. This is
the classical architecture of the early Internet, where
users were often offline and so servers were necessary.
Open standards for protocols usually have this archi-
tecture, and so clients are not restricted to a single
server, as otherwise server-mediated communication
between multiple clients on different servers would be
impossible. Taking e-mail for example, the user oper-
ates a client known as the Mail User Agent (MUA)
that sends mail via one or more the Mail Transfer
Agents (MTA, i.e. “email servers”) until reaching the
recipient’s server. This server in turn sends the mail to
the recipient’s MUA. The security of the e-mail eco-
system at least assumes each of these channels is en-
crypted via TLS (Foster et al., 2015). SMTP services
such as Gmail are federated, as would be any ser-
vice using a standardized messaging protocol such as
XMPP+OTR or even “federated” or “permissioned”
blockchains.
4.3 Decentralized
In the decentralized model also a “peer to peer”
(P2P) model) clients may directly communicate
without a server, and so clients are considered
“peers” (Minar and Hedlund, 2001). In a peer-to-
peer framework, clients may pass messages on behalf
of other clients, although they may also drop mes-
sages. Servers are not used. This naturally leads to
a default to broadcast messaging in peer-to-peer sys-
tems. Earlier systems like Bittorrent as well as newer
“permission-less” blockchain-based models fall into
this category.
A proper subset of distributed systems that oper-
ate in an adversarial model are decentralized systems,
which are “systems in which multiple authorities con-
trol different components and no single authority is
fully trusted by all components” (Troncoso et al.,
2017) As per malicious security (Yung, 2015), both
centralized and decentralized parties may no longer
honestly follow the protocol, but may maliciously at-
tempt to subvert the protocol: For example, a peer in a
peer-to-peer system may drop traffic or claim falsely
to execute part of the protocol while not doing so.
Distributed ledgers, and implementations such as
blockchains, allow components of a system to record
their activity transparently so that other peers can ver-
ify if components are following the protocol and ma-
licious components detected (Halpin and Piekarska,
2017). Distributed consensus protocols are then re-
quired to when the components need to have a consis-
tent global view of some part of the system (Lamport
et al., 1982), such as a distributed ledger of transac-
tions. Distributed systems that do not require the quo-
rum be set in advance but allow any peer to join are
permissionless, and so decentralized. In contrast, if a
quorum of existing members or other manual proce-
dure is required to allow a new peer to enter the sys-
tem, then the manual procedure is a centralized root
of trust or the quorum of “trusted” peers equivalent to
servers in the federated model in practice.
5 ANALYSIS
The decentralization trilemma
4
states that a system
can only have two of three in terms of security, scal-
ability, or decentralization but not all three. On one
hand, a system may be both secure and scalable, but
not decentralized. Google and other cloud-based ser-
vices fit in this requirement. In contrast, a system may
be decentralized and scalable, but not secure. Systems
such as Bittorrent fall in this category (Wolchok et al.,
2010). Finally, a system may be decentralized and se-
cure, but not scalable. Blockchains such as Ethereum
and Bittorrent would be part of this category.
Federated systems attempts to balance the
trilemma by having multiple servers, but in reality
this simply spreads the trust from one server to mul-
tiple servers. Therefore, federated systems are not
“trustless” and so not decentralized, even if they use a
blockchain. Some of the examples are subtle: Napster
did use a peer-to-peer protocol, but had a centralized
search directory. So Napster would count as feder-
ated, as the centralized directory server was required
for the function of searching torrents, and each peer
needs to connect to that server in order to function
properly.
By tying the decentralization trilemma to con-
crete technical properties, we can determine if it ac-
tually holds. Every architecture makes certain design
4
https://github.com/ethereum/wiki/wiki/Sharding-FAQ
SECRYPT 2020 - 17th International Conference on Security and Cryptography
508
choices that privilege one property over another. Al-
though more work is necessary in order to formally
characterize each of these architectures, history sug-
gests that these architectures are structurally biased
toward certain properties and against others. Table 1
provides a comparison of the choices made by the ar-
chitectures outlined above.
5.1 Security Analysis
5.1.1 Confidentiality, Integrity, and Authenticity
In general, all architectures are able to imple-
ment classical security properties using cryptography.
However, authentication is a problem also for all ar-
chitectures, albeit in different ways. Centralized ar-
chitectures have low authenticity insofar as a user can
only communicate to other users via a server and a
server could serve false keys or impersonate its users,
as is possible even with Signal (Kobeissi et al., 2017).
This critique also applies to federated servers, al-
though out-of-band finger-print verification may help,
despite users often being unable to use fingerprint ver-
ification to detect a man-in-the-middle attack by a ma-
licious server (Schr
¨
oder et al., 2016). Peer to peer sys-
tems suffer from similar issues, as key discovery can
be easily manipulated by a malicious peer. However,
decentralization at least allows direct communication
of key material and non-custodial key ownership, and
so decentralized systems have an advantage as a (pos-
sibly malicious) server is not required to mediate key
material. Also, decentralization requiring each client
to serve as a peer leads to an inability to cover certain
use-cases, as security properties need to be enforced
over groups, but each peer is meant to be an individ-
ual. Security operations over groups are more well-
defined but much less commonly deployed (Chaum
and van Heyst, 1991). For example, Bitcoin has issues
with large group signatures and secure group messag-
ing is still considered a hard problem, even in central-
ized environments (Goldberg et al., 2009).
5.1.2 Privacy: Unlinkability and
Unobservability
Without advanced techniques like differential privacy
or secure multi-party computation, centralization has
low unlinkability insofar as the malicious server can
link any of its users to messages and can always map
the network of a user via observing their messages.
Likewise for authenticity issues in centralized sys-
tems, and the same is possible in federated settings.
Decentralized systems by design can offer possibly
more privacy as it is less trivial for a central server
to monitor all transactions, although not without cost.
For example, even Tor is vulnerable to traffic analy-
sis attacks on the exit and entrance nodes. (Dingledine
et al., 2004). Some decentralized designs allow a high
amount of unlinkability, using techniques like zero-
knowledge proofs such as ZCash (Ben-Sasson et al.,
2014) and unobservability via the use of cover traf-
fic (Clarke et al., 2000), although these designs are
not widely deployed. The problem is cover traffic
and other techniques to increase anonymity come at
the cost of reducing the capacity of the network, as
given by the well-known “anonymity trilemma” be-
tween capacity, latency, and anonymity (Das et al.,
2018).
5.2 Scalability Analysis
5.2.1 Availability, Capacity, and Latency
Scalability concerns the reliable operations of the sys-
tem in the face of growing demand. Due to their use
of robust distributed (but not decentralized) systems,
centralized architectures are highly available and can
easily scale to having a high capacity with low la-
tency (DeCandia et al., 2007). Federated systems
have as a bottle-neck the server with the least capacity
(similar to decentralized systems). In most decentral-
ized architectures, availability is low as peers may al-
ways be offline. Although federation may be a disad-
vantage in terms of unlinkability in comparison to de-
centralized systems, federated systems typically have
much higher availability than decentralized designs as
the server is normally online while peer-to-peer sys-
tems suffer from network churn. This tends to trans-
late into lower latency in comparison to decentralized
systems. In centralized systems, it is much easier to
upgrade capacity via simply upgrading servers and
connection capacity. In decentralized systems, the
problem of the “weakest” link of capacity can be pro-
found, particularly if there is no way to route around
low capacity peers. Therefore, real-world decentral-
ized systems like Bittorrent tend to evolve “super-
nodes” that have high capacity and are online. These
“super nodes” in effect emerge as de-facto “servers”
in the decentralized system (Wang and Kangasharju,
2013), undermining the security and privacy proper-
ties of decentralized systems and making them de-
facto federated systems.
5.2.2 Transparency and Portability
Centralized and federated servers whose code and op-
erations are hidden from their users on a server are
difficult for third parties to audit by their users, and
so transparency is low. Decentralized systems offer
less ability for an adversary to tamper with incoming
Deconstructing the Decentralization Trilemma
509
Table 1: Properties compared to architectures (* assumes a distributed ledger).
Property Centralized Federated Decentralized
Security and Privacy Properties
Confidentiality 3 3 3
Integrity 3 3 3
Authenticity 7 7 3
Unlinkability 7 7 3
Unobservability 7 7 7
Scalability Properties
Availability 3 3 7
Capacity 3 7 7
Latency 3 3 7
Transparency 7 7 3*
Portability 7 3 7
Examples: Apple Pay, Signal Visa, E-mail (PGP) Bitcoin, Bittorrent
messages as the malicious server has been eliminated
insofar as messages may be directly sent from user
to user, although in practice a relaying peer may tam-
per with the messages. For example, Bittorrent has no
transparent log, and so nodes dropping packets cannot
be easily detected.
Although decentralization gets rid of a malicious
server, clients may join and act maliciously in a sybil
attack (Douceur, 2002). Therefore, blockchain tech-
nology requires enforcing transparency in a decentral-
ized setting so attacks and failures by peers can be
discovered. In terms of portability, decentralized sys-
tems suffer from a dearth of interoperable standards,
while federated systems are highly standardized (e-
mail, XMPP, etc.). The usability of most peer-to-peer
systems tend to be poor, while federated systems such
as email have been able to compete in terms of usabil-
ity with centralized systems.
6 OPEN PROBLEMS
Our analysis leads to a number of open problems
exist that prevent the further development of decen-
tralized alternatives. Some of the disadvantages of
decentralization can be solved by standardization in
order to bring the portability of decentralized archi-
tectures in competition with well-known federated
protocols. Better usability is desperately needed to
drive more users from centralized to decentralized
systems. Some of the characteristics of architectures
could be changed by technical advances: For exam-
ple, increased usage of cover traffic could help ad-
dress unobservability when combined with mix net-
working (Danezis et al., 2010). Yet some problems
are more fundamental. In surveys of secure communi-
cation (Unger et al., 2015), a pattern starts to emerge:
Any serious attempt to build a decentralized alterna-
tive that is resistant to a malicious server eventually
comes up against similar hard problems.
It is possible to ignore many of these problems if
one rules out one of these properties, but users have
grown accustomed to contemporary methods of on-
line communication based on silos (Facebook, Twit-
ter, etc.) that are easy to use and have high availabil-
ity, as well as assume a level of authenticity that is
normally not justified by the difficulties of trusted key
discovery in centralized systems. So, decentralized
architectures that are designed to defend against ma-
licious servers should find solutions to the “seven hard
problems” as follows:
5
1. Public Key Discovery Problem: Public key dis-
covery and validation is very difficult for users to
manage in a decentralized system, but without it a
message cannot have authenticity.
2. Key Availability Problem: Users would like
communicate seamlessly and securely across dif-
ferent devices, as well as restore data if a device
or key is lost.
3. Group Problem: Users work in social groups,
yet public key cryptography does not have a con-
sistent manner for dealing with groups in decen-
tralized settings. This effects confidentiality, in-
tegrity, and authenticity.
4. Metadata Problem: Messages are vulnerable to
traffic analysis, unless cover traffic is used, which
is not common in decentralized systems. This de-
rives from issues around unlinkability and unob-
servability.
5
An earlier version of this list was originally defined at
LEAP: https://leap.se/en/docs/tech/hard-problems
SECRYPT 2020 - 17th International Conference on Security and Cryptography
510
5. Sybil Problem: Without servers, any client may
join a decentralized network and act maliciously
in a sybil attack (Douceur, 2002).
6. Update Problem: Software updates need to be
securely delivered across multiple clients in a de-
centralized system, which is easy to do with a cen-
tralized server.
7. Resource Problem: It is an open problem on how
to let users securely share a resource for real-time
collaboration in a decentralized architecture.
These problems cause real-world impact on inter-
est in decentralized architectures both by users and
researchers. New standards like IETF Message Layer
Security seem to be allowing encrypted user commu-
nication to tackle the group problem, but it is not clear
how to apply to decentralized systems.
6
The update
problem is partially addressed in terms of security by
reproducible builds. The resource problem in part is
due to the group problem being unsolved in tandem
with scalability issues, where decentralized designs
have a lack of availability and latency that make shar-
ing a single resource between a group difficult.
7 CONCLUSIONS
The decentralization trilemma is a simplification that
can be deconstructed into traditional security and
scalability properties. Nonetheless, decentralization
solves a real problem: Decentralization prevents a
single centralized malicious server from compromis-
ing the security of users. There is a kernel of truth
in the decentralization trilemma: In broad strokes,
decentralization does offer better security at the cost
of scalability if end-users manage their keys and can
offer higher scalability if sacrificing security is ac-
ceptable. Nonetheless, decentralized systems natu-
rally have scaling issues as nodes are more frequently
off-line or faulty. Worse, communication can require
more hops than in centralized systems, increasing the
chance of failure.
The real breakthrough in blockchain technology,
in comparison to earlier peer-to-peer decentralized
systems, is the use of distributed ledgers to prevent
malicious behavior by peers via transparency. This
naturally leads to loss of latency due to the nature
of distributed consensus protocols (Lamport, 1978).
Blockchain technology, while not solving the fun-
damental issues around availability that plague de-
centralized systems, does allow malicious servers to
be removed and malicious peers to be detected, and
6
https://datatracker.ietf.org/wg/mls/
could lessen the security problems caused by “super
nodes. If peers could be ubiquitously online, they
should be more available, which may increase capac-
ity and decrease latency. Yet to do so securely would
mean the network would reach widespread usage and
users would be highly security-conscious, which ap-
pears delusional as most users are not systems admin-
istrators.
The low availability of decentralized architectures
likely necessitates learning lessons about the evolu-
tion of “super-nodes” (high capacity nodes that relay
the majority of the traffic) in peer-to-peer architec-
tures, which can act as high availability relays and so
bring the benefits of federated designs to decentral-
ized systems. In order to secure these “super-nodes,
it does seem to make sense that a specialized class of
system administrators would arise in federated sys-
tems.
Thus, the technical problem of decentralization is
revealed at its core to be a political problem: Central-
ization to a large extent arises due to the monopoly
of technical knowledge by a class of system admin-
istrators and programmers. The goal of decentralized
systems seems to be to spread this technical knowl-
edge so that all users can autonomously operate and
govern their own infrastructure. In this way, federated
and centralized models represent the inherent trust as-
sumptions of non-technical users. Historically, hu-
mans typically trust friends and associates, as well
as deploy a specialization of labor. So a user would
likely trust a highly-skilled individual they know in
some fashion, perhaps a friend or affiliate at a human-
scale institution like a university, to manage technical
infrastructure on their behalf. The rise of a non-profit
federated systems at the dawn of the Internet maps to
pre-existing human trust relationships. As the techni-
cal complexity of the Internet scaled and users needed
more convenience, this model shifted as users trusted
Google to handle the sheer technical complexity in re-
turn for profit off of personal data or for a certain cost.
The lack of trust in these centralized servers and the
humans behind them is what led to the rise of peer-
to-peer and blockchain: Decentralization is a political
ideology masquerading as a technical architecture.
The next steps to explore this thesis would re-
quire formalizing properties such as availability, la-
tency, and capacity in a more rigorous manner than
presented here, similar as what has been done in re-
search on privacy (Das et al., 2018). Then archi-
tectures could be compared with the same rigor as
security properties within a formal message-passing
framework as done by the π-calculus (Kobeissi et al.,
2017). Another promising avenue is using network-
theoretic approaches to study the evolution of decen-
Deconstructing the Decentralization Trilemma
511
tralized networks into super-nodes. Further socio-
technical work in needed in understanding the mo-
tivations, usability, and political stakes of decentral-
ization. The ideology of decentralization has given us
very real advances in technology for ordinary users,
but much more is needed for liberation.
REFERENCES
Baran, P. et al. (1964). On distributed communications. Vol-
umes I-XI, RAND Corporation Research Documents,
August.
Ben-Sasson, E., Chiesa, A., Garman, C., Green, M., Miers,
I., Tromer, E., and Virza, M. (2014). Zerocash: De-
centralized anonymous payments from bitcoin. In
IEEE Symposium on Security and Privacy.
Chaum, D. and van Heyst, E. (1991). Group signatures. In
Advances in Cryptology.
Clarke, I., Sandberg, O., Wiley, B., and Hong, T. W. (2000).
Freenet: A distributed anonymous information stor-
age and retrieval system. In Designing Privacy En-
hancing Technologies, International Workshop on De-
sign Issues in Anonymity and Unobservability.
Danezis, G., D
´
ıaz, C., Troncoso, C., and Laurie, B. (2010).
Drac: An architecture for anonymous low-volume
communications. In 10th Privacy Enhancing Tech-
nologies Symposium.
Das, D., Meiser, S., Mohammadi, E., and Kate, A. (2018).
Anonymity trilemma: Strong anonymity, low band-
width overhead, low latency-choose two. In 2018
IEEE Symposium on Security and Privacy (SP), pages
108–126. IEEE.
DeCandia, G., Hastorun, D., Jampani, M., Kakulapati,
G., Lakshman, A., Pilchin, A., Sivasubramanian, S.,
Vosshall, P., and Vogels, W. (2007). Dynamo: Ama-
zon’s highly available key-value store. In ACM Sym-
posium on Operating Systems Principles: Proceed-
ings of twenty-first ACM SIGOPS symposium on Op-
erating systems principles, volume 14, pages 205–
220.
Dingledine, R., Mathewson, N., and Syverson, P. (2004).
Tor: The second-generation onion router. Proceedings
of the 13th USENIX Security Symposium, 2.
Douceur, J. R. (2002). The sybil attack. In 1st International
Worksop on Peer-to-Peer Systems.
Foster, I. D., Larson, J., Masich, M., Snoeren, A. C., Sav-
age, S., and Levchenko, K. (2015). Security by any
other name: On the effectiveness of provider based
email security. In Proceedings of the 22nd ACM
SIGSAC Conference on Computer and Communica-
tions Security, pages 450–464. ACM.
Goldberg, I., Ustaoglu, B., Gundy, M. V., and Chen, H.
(2009). Multi-party off-the-record messaging. In 16th
ACM Conference on Computer and Communications
Security.
Halpin, H. and Piekarska, M. (2017). Introduction to se-
curity and privacy on the blockchain. In 2017 IEEE
European Symposium on Security and Privacy Work-
shops (EuroS&PW), pages 1–3. IEEE.
Kobeissi, N., Bhargavan, K., and Blanchet, B. (2017). Auto-
mated verification for secure messaging protocols and
their implementations: A symbolic and computational
approach. In 2017 IEEE European Symposium on Se-
curity and Privacy (EuroS&P), pages 435–450. IEEE.
Lamport, L. (1978). Time, clocks, and the ordering of
events in a distributed system. Communications of the
ACM, 21(7).
Lamport, L., Shostak, R., and Pease, M. (1982). The
byzantine generals problem. ACM Transactions on
Programming Languages and Systems (TOPLAS),
4(3):382–401.
Minar, N. and Hedlund, P. (2001). A network of peers
peer-to-peer models through the history of the in-
ternet. In Oram, A., editor, Peer-to-Peer: Harness-
ing the Power of Disruptive Technologies, pages 9–20.
O’Reilly, Sebastopol, CA.
Pfitzmann, A. and Hansen, M. (2005). Anonymity, un-
linkability, unobservability, pseudonymity, and iden-
tity management a consolidated proposal for termi-
nology. Technical report.
Schr
¨
oder, S., Huber, M., Wind, D., and Rottermanner, C.
(2016). When signal hits the fan: on the usability and
security of state-of-the-art secure mobile messaging.
In European Workshop on Usable Security. IEEE.
Slobogin, C. (2014). Cause to believe what?: The impor-
tance of defining a search’s object–Or, how the ABA
would analyze the NSA metadata surveillance pro-
gram. Oklahoma Law Review, pages 14–7.
Troncoso, C., Isaakidis, M., Danezis, G., and Halpin,
H. (2017). Systematizing decentralization and pri-
vacy: Lessons from 15 years of research and deploy-
ments. Proceedings on Privacy Enhancing Technolo-
gies, 2017(4):404–426.
Unger, N., Dechand, S., Bonneau, J., Fahl, S., Perl, H.,
Goldberg, I., and Smith, M. (2015). Sok: Secure mes-
saging. In 2015 IEEE Symposium on Security and Pri-
vacy, pages 232–249. IEEE.
Wang, L. and Kangasharju, J. (2013). Measuring large-
scale distributed systems: case of BitTorrent mainline
DHT. In 13th IEEE International Conference on Peer-
to-Peer Computing.
Wolchok, S., Hofmann, O. S., Heninger, N., Felten, E. W.,
Halderman, J. A., Rossbach, C. J., Waters, B., and
Witchel, E. (2010). Defeating vanish with low-cost
sybil attacks against large dhts. In Network and Dis-
tributed System Security Symposium.
Yung, M. (2015). The mobile adversary paradigm in dis-
tributed computation and systems. In Proceedings
of the 2015 ACM Symposium on Principles of Dis-
tributed Computing, pages 171–172. ACM.
SECRYPT 2020 - 17th International Conference on Security and Cryptography
512