Efficient Joint Random Number Generation for Secure Multi-party
Computation
Erwin Hoogerwerf
1
, Daphne van Tetering
1
, Aslı Bay
2
and Zekeriya Erkin
1,3
1
Cyber Security Group, Delft University of Technology, The Netherlands
2
Department of Computer Engineering, Antalya Bilim University, Antalya, Turkey
3
Digital Security Group, Radboud University Nijmegen, The Netherlands
Keywords:
Joint Random Number Generation, Secure Multi-party Computation, Data Aggregation.
Abstract:
Large availability of smart devices and an increased number of online activities result in extensive personalized
or customized services in many domains. However, the data these services mostly rely on are highly privacy-
sensitive, as in pace-makers. In the last decades, many privacy breaches have increased privacy awareness,
leading to stricter regulations on data processing. To comply with this legislation, proper privacy preserva-
tion mechanisms are required. One of the technological solutions, which is also provably secure, is Secure
Multi-Party Computation (SMPC) that can compute any function with secret inputs. Mainly, in several SMPC
solutions, such as data aggregation, we observe that secret values distributed among parties are masked with
random numbers, encrypted and combined to yield the desired outcome. To ensure correct decryption of the
final result, it is required that these numbers sum to a publicly known value, for instance, zero. Despite its
importance, many of the corresponding works omit how to obtain such random numbers jointly or suggest
procedures with high computational and communication overhead. This paper proposes two novel protocols
for Joint Random Number Generation with very low computational and communication overhead. Our pro-
tocols are stand-alone and not embedded in others, and can therefore be used in data aggregation and other
applications, for instance, machine learning algorithms, that require such random numbers. We first propose
a protocol that relies on bit-wise sharing of individually generated random numbers, allowing parties to adapt
random numbers to yield a public sum. Second, we propose a protocol that uses the sign function to generate
a random number from broadcast numbers. We provide security and complexity analyses of our protocols.
1 INTRODUCTION
Recent developments in machine learning and data
analytics have allowed more and more customized
personalized online services, thanks to the availability
of vast numbers and types of data collected all around
us. On the one hand, the data collected can be used
to improve machine learning algorithms and provide
users with personalized services such as news noti-
fications or product recommendations. On the other
hand, data collection raises serious privacy concerns
since often sensitive data is included (Jiang et al.,
2019; Zhao et al., 2019). Over the last years, we have
witnessed numerous privacy breaches (Privacy Inter-
national, 2020), damaging individuals and also busi-
nesses. These privacy breaches have led to increased
privacy awareness, not only among users but also in
the political arena, resulting in stricter laws on data
processing such as the General Data Protection Regu-
lation in Europe and the Personal Data Protection Act
in Singapore (European Parliament and Council of
European Union, 2016; Singapore Parliament, 2014).
To comply with legislation, proper privacy preserva-
tion mechanisms are required. Among many other
technological solutions, Secure Multi-Party Compu-
tation (SMPC) techniques attract more attention from
academia and business as it provides provable secu-
rity in different adversarial models.
Firstly introduced by Yao, SMPC protocols are
designed to evaluate a function using secret inputs
distributed among different parties, thereby yielding
a public result (Yao, 1982). Secure Multi-Party Com-
putation is used to perform, among others, data clus-
tering (Erkin et al., 2013), homomorphic encryp-
tion (Lyu, 2020) and data aggregation (Erkin and
Tsudik, 2012). In this paper, we focus on a particular
computation, namely Joint Random Number Genera-
tion (JRNG), which is notably used a lot in privacy-
436
Hoogerwerf, E., van Tetering, D., Bay, A. and Erkin, Z.
Efficient Joint Random Number Generation for Secure Multi-party Computation.
DOI: 10.5220/0010534804360443
In Proceedings of the 18th International Conference on Security and Cryptography (SECRYPT 2021), pages 436-443
ISBN: 978-989-758-524-1
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
preserving data aggregation protocols. In smart me-
ter settings, for example, individual measurements are
sent to a data aggregator, responsible for computing
aggregate statistics such as sum or average. Previous
research has shown that these measurements, when
obtained by adversaries in an attack on data confi-
dentiality, can be used to infer patterns revealing cus-
tomer behaviour (Wang and Lu, 2013). To prevent
such eavesdropping attacks from happening, smart
meters are required to encrypt their measurements be-
fore sharing these with the aggregator. However, en-
cryption is not sufficient to prevent malicious aggre-
gators from inferring additional knowledge (Kursawe
et al., 2011). One way to achieve this and protect
against malicious aggregators is by requiring all me-
ters to mask their measurements with random num-
bers. Jointly generating these among all participants
in the protocol ensures they sum to a publicly known
value, thereby guaranteeing the final result’s correct
decryption. Therefore, Joint Random Number Gener-
ation can be considered an important step in privacy-
preserving data aggregation protocols.
Remarkably, most research on privacy-preserving
data aggregation omits how to obtain such numbers
and assumes their existence. For example, Shi et
al. assume that all keys during the setup of the pri-
vate stream aggregation algorithm sum to zero (Shi
et al., 2011; Shi et al., 2015). Lu et al. assume the
same in their homomorphic encryption protocol (Lu
et al., 2017). Moreover, several studies that do fo-
cus on JRNG have high computational and communi-
cation overhead. For example, Erkin and Tsudik as-
sume a fully connected network is available, allowing
each party to exchange random values with all others
(Erkin and Tsudik, 2012). With large networks, this
approach becomes infeasible. A similar approach by
Kursawe et al. uses the notion of leaders, who are re-
quired to compute their random values such that all
values together sum to zero (Kursawe et al., 2011).
This reduces the complexity for all other nodes in the
network since these only communicate with the set
of leaders. However, the complexity of leader nodes
remains the same as in (Erkin and Tsudik, 2012).
This paper presents two novel protocols to per-
form Joint Random Number Generation with minimal
computational and communication overhead. The
first protocol relies on bit-wise sharing of individu-
ally generated random numbers. By doing so, for
each round, the information leaked to an adversary
is minimized to only one bit per participant. The
second protocol uses a single broadcast and is based
on Diffie-Hellman key exchange (Merkle, 1978). We
also present a version of the broadcast-based proto-
col that has reduced complexity. Finally, to compare
all protocols, we conduct an analysis considering both
communication and computation complexity.
Our contributions are two-fold. First, by formaliz-
ing two JRNG protocols explicitly, we provide clarity
on how to efficiently perform Joint Random Number
Generation. In contrast to previous work, our pro-
tocols are stand-alone and not embedded in others.
Therefore, they can be used in any domain or applica-
tion where JRNG is required. Finally, we show how to
perform JRNG with minimal computational and com-
munication overhead, thus improving the way JRNG
is done for the whole research community.
The remainder of this paper is structured as fol-
lows. First, we discuss related work and preliminar-
ies in section 2 and section 3. Next, we discuss the
share-based protocol in section 4, before discussing
the broadcast-based protocol in section 5, along with
its security proof. Finally, discuss complexity in sec-
tion 6 and conclude the paper in section 7.
2 RELATED WORK
Related work on Joint Random Number Generation
(JRNG) can be divided into two categories: research
that focuses on JRNG as a main goal, as does this
one, and research that uses it to achieve a broader
goal, and therefore embeds it in its protocol. First,
we discuss work that embeds JRNG to, for example,
perform privacy-preserving data aggregation: jointly
generated random numbers are used to mask indi-
vidual values before sending them to the aggregator.
When the aggregator sums or multiplies all values,
the masks cancel out and allow the aggregator to ob-
tain the exact result without revealing individual val-
ues. Since in this work, we care about JRNG specif-
ically, we only highlight those procedures and omit
any other details. An example of such research is the
work by Erkin and Tsudik, performing additive ag-
gregation of plaintexts in a setting with at least three
parties (K 3) (Erkin and Tsudik, 2012). First, par-
ties exchange random numbers in a pair-wise man-
ner, i.e. parties p
i
and p
j
exchange values r
(p
i
p
j
)
and
r
(p
j
p
i
)
. Then, each party computes its random value
R
p
i
by summing over the K 1 received values:
R
p
i
=
K
j=1, j6=i
r
(p
i
p
j
)
r
(p
j
p
i
)
. (1)
It is trivial to see that the addition of these elements
yields zero. While presented in a smart meter set-
ting, this scheme can be used in any setting where
JRNG is required. However, it should be noted that
this work requires the presence of secure communica-
tion channels. The second work is an interactive pro-
tocol proposed by Kursawe, Danezis, and Kohlweiss
Efficient Joint Random Number Generation for Secure Multi-party Computation
437
(Kursawe et al., 2011). Similar to Erkin and Tsudik,
it jointly generates random numbers yielding a zero-
sum. However, to reduce communication complexity,
it uses the notion of leaders. At the start of the proto-
col, each party generates a random number and shares
it with the set of leaders P. Then, each leader gener-
ates its random number in such a way that summing it
with all received shares yields zero. Similar to Erkin
and Tsudik, this protocol requires at least three parties
and secure communication channels. The use of lead-
ers reduces communication complexity for non-leader
parties, the complexity for leaders is unchanged. This
protocol is adapted by Van de Kamp et al. to remove
the notion of leaders: instead, parties share their ran-
dom number with all other parties (Van De Kamp
et al., 2019). In the same work, Kursawe et al. propose
a JRNG scheme based on the Diffie-Hellman key ex-
change. The scheme yields a publicly known product
rather than a sum, using the sign function s
i
( j).
s
i
( j) =
(
1 if i > j ,
1 if i < j .
(2)
In this protocol, each party broadcasts its public key
g
k
p
i
, for which g Z
p
is a public generator and k
p
i
Z
p
is a random secret key. Each party uses these num-
bers to compute its random number R
p
i
as follows:
R
p
i
=
K
j=1, j6=i
(g
k
p
j
)
s
i
( j)k
p
i
= g
K
j=1, j6=i
s
i
( j)k
p
i
k
p
j
. (3)
Verifying that this yields one is done by
K
i=1
R
p
i
= g
K
i=1
K
j=1, j6=i
s
i
( j)k
p
i
k
p
j
= g
K
i=1
i1
j=1
(s
i
( j)+s
j
(i))k
p
i
k
p
j
= g
0
= 1 .
When using this protocol for aggregation, each party
masks its individual value c
i, j
using randomness R
p
i
and sends this to the aggregator as g
c
i, j
+R
p
i
. As veri-
fied above, all random numbers cancel out in the final
sum, leaving the aggregator with g
i
c
i, j
. Since it is not
trivial to find the correct final sum from this, Kursawe
et al. argue that the aggregator should launch a brute-
force attack to obtain it (Kursawe et al., 2011). More
recently, Bell et al. embedded JRNG in an aggrega-
tion protocol by requiring parties to establish shared
keys with their neighbours to derive pairwise random
masks, masking input c
i
as follows (Bell et al., 2020):
c
i
j<i
m
i, j
+
j>i
m
i, j
. (4)
This method again ensures that all masks cancel out
in the final sum. The protocol operates in a semi-
honest security setting. It can be extended to a so-
called semi-malicious security setting, meaning that
the aggregation server is required to perform the first
step of the protocol honestly, but is allowed to deviate
from the protocol after that (Bell et al., 2020).
Finally, we close this section by discussing work
that specifically focuses on Joint Random Number
Generation. In this work, JRNG is denoted by its au-
thors as modulo zero-sum randomness (Hayashi and
Koshiba, 2018). As the name explains, modulo zero-
sum randomness means that when summing individ-
ual numbers, the result is zero. To ensure zero-sum
randomness, three conditions have to be met:
1. Modulo Zero Condition: the sum of all random
numbers X
i
has to be zero.
2. Independence Condition: all numbers X
1
,...,X
m
used have to be generated independently.
3. Secrecy Condition: each player i has knowledge
of the random number it generated itself, X
i
, and
of the fact, the sum of all numbers will yield zero,
but has no additional knowledge.
In this work, to generate random numbers, the authors
impose a cyclic ordering on all parties. First, each
party agrees on a secret random number with its two
neighbors, Z
i
and Z
i1
. Then, parties determine their
final random by subtracting these two values:
R
p
i
= Z
i
Z
i1
. (5)
It is trivial to see that in summation, this yields zero.
In follow-up work, the use of multiple access chan-
nels and verifiable quantum protocols to generate
modulo zero-sum randomness is discussed (Hayashi,
2018; Hayashi and Koshiba, 2019).
It is important to note that all related work in this
section focuses on yielding a public sum of zero. Our
approach is a more generalized one, where instead we
facilitate the joint generation of random numbers that
sum up to a publicly known value N 0.
3 PRELIMINARIES
In this section, we introduce preliminaries, the secu-
rity assumption, and notation used in this work.
3.1 Decisional Diffie-Hellman Problem
The security of the broadcast-based scheme presented
in section 5 relies on the Decisional Diffie-Hellman
Problem (DDH). This problem is defined as follows:
given a cyclic group G of prime order q, its public
generator g, the following two distinguishers are com-
putationally indistinguishable (in the security param-
eter q ); (g
x
,g
y
,g
xy
), where x,y are chosen uniformly at
random from [0,... q 1] and (g
x
,g
y
,g
z
), where x,y,z
are chosen uniformly at random from [0,... q 1].
SECRYPT 2021 - 18th International Conference on Security and Cryptography
438
3.2 Security Setting
We assume an environment composed of a static
group of K entities wanting to jointly establish ran-
dom numbers with a publicly known sum. Similar
to related work, we assume all entities to be semi-
honest (honest-but-curious). This means they follow
the protocol but are not prevented from inferring ad-
ditional knowledge from received information. Each
entity is capable of performing modular operations,
such as modular addition, and of generating strong
random numbers. We do not assume the presence of
any other active entity, such as a trusted third party or
an aggregator, to perform computations. Protecting
against active adversaries can be done at the cost of
additional communication and computation complex-
ity, but is considered out of the scope of this work.
For the bit-wise sharing protocol, we assume the
existence of secure communication channels. These
ensure that any message sent over the network can
only be retrieved by its sender and recipient.
3.3 Notation
The notation used is summarized in Table 1.
Table 1: Notation Summary.
Symbol Definition
p large prime
r
p
i
p
j
random value generated by p
i
send to p
j
g generator for Z
p
K number of entities in the protocol
l security parameter
m bit-length of random numbers
C
i
carry-over from bit i
N publicly known target value
R
p
i
random number generated by entity p
i
k
p
i
temporary values used to construct R
p
i
x
i j
bit j of random number R
p
i
x
N j
bit j of publicly known target value N
y
i
updated bit i replacing x
ii
of R
i
d
i
scaled carry-over for key entity i
s
i
( j) sign function
Z
a
short for Z/aZ
4 SHARE-BASED JRNG
PROTOCOL
In this section, we describe a Joint Random Number
Generation protocol that uses the bit-wise sharing of
individually generated random numbers. The random
numbers generated by the protocol sum up to a pub-
licly known integer 0 N 2
m
, for some m N. Ini-
tially, all entities generate a random number R
p
i
, and
a subset of all entities is selected as m key entities.
Then, the acting key entity adjusts the j’th bit of its
random number such that it equals the j’th bit of the
target N. To do so, each entity sends its jth bit to
the acting key entity, which sums these to determine
if its j’th bit should change. Finally, to prevent carry-
over from affecting subsequent computations, the act-
ing key entity subtracts this from its random number.
This process is repeated for m iterations until each key
entity has adjusted its random number once.
The protocol is formalized as follows:
1. All K entities generate random numbers R
p
i
Z
2
m
, such that R
p
i
=
m
j=1
x
i j
2
j1
,x
i j
{0,1}.
m entities are assigned key entity as in (Kursawe
et al., 2011).
2. Then, for m iterations:
2.1. All entities i send x
i j
to the acting key entity
using the underlying secure network.
2.2. The acting key entity calculates d
j
=
K
i=1,i6= j
x
i j
. If d
j
x
N j
mod 2, it sets its
j’th coefficient to y
j
= 0. Otherwise, it sets its
j’th coefficient to y
j
= 1.
2.3. Finally, the acting key entity calculates C
j
=
b
d
j
+y
j
2
c2
j
and updates its random number to
R
j
R
j
C
j
mod 2
m
.
In practice, usually N = 0. However, to keep the final
outcome secret, different values for N can be used.
Then, the final outcome can only be determined by the
entities that know N, since these can subtract it from
the final outcome. While in this work the protocol is
demonstrated in a binary setting, it can be used in any
base-k setting for some k N.
A distinguishing feature of this protocol is that in-
stead of communicating full-sized random numbers
in each round, this protocol uses bit-wise communi-
cation. This reduces communication complexity to a
total of K 1 bits per iteration. This reduction be-
comes more significant for large numbers and high
amounts of entities. Since only key entities are re-
quired to perform modular operations, the scheme can
be used in situations where not all entities have equal
or sufficient computing power. Even more important,
the use of key entities ensures that computation com-
plexity is dominated by the number of key entities,
not by the number of entities in general. Compared
to the work by Erkin and Tsudik, (Erkin and Tsudik,
2012), communication complexity is reduced and a
fully connected network is not required.
It is important to note that this scheme should not
be used for applications where certain bits are more
Efficient Joint Random Number Generation for Secure Multi-party Computation
439
relevant than others, as is the case in a bank transfer.
If this protocol would be used to mask a bank transfer
with encryption c = m +R mod 2
m
, this would allow
the adversary to distinguish between high-valued and
low-valued transactions by only corrupting one entity.
Finally, it should be noted that the scheme can also
be used for situations where m > K. This can be done
as follows: for each j, where K < j m, we require
entity K to act during all remaining iterations, thereby
continuously adjusting its random number.
4.1 Security
When assessing the share-based protocol in a semi-
honest security setting, it becomes clear that the pro-
tocol is vulnerable to collusion among key entities. In
each round, the acting key entity receives an entity’s
j’th bit, thereby obtaining partial information about
each random number R
p
i
. Because of this, collusion
allows key entities to partially reconstruct the ran-
dom number held by other entities. While this seem-
ingly renders the protocol unusable, situations exist
in which computational speed is more important and
where non-privacy-preserving Joint Random Number
Generation suffices. Such a situation is found in the
work by Garay, Schoenmakers, and Villegas (Garay
et al., 2007). Here, a random number of r is jointly
generated, before being encrypted using fresh ran-
domness. By adding fresh randomness to r, it is ef-
fectively masked. Therefore, in this case, our bit-wise
sharing protocol provides sufficient security.
In situations where privacy of individual bits is
required, secure multi-party computation techniques,
such as (additive) secret sharing can be added to the
protocol (Van De Kamp et al., 2019). This can be
achieved as follows: during the second step of the
protocol, instead of directly sending its bit x
i j
to the
acting key entity, each entity first generates K 1
shares s
i j
of its bit, such that x
i j
=
K1
j=1
s
i j
mod N,
similar to the work by Garcia and Jacobs (Garcia and
Jacobs, 2010). Of these, it holds one by itself and
distributes the others among other non-acting key en-
tities. Finally, instead of sending entire bits, each en-
tity sends the received shares to the acting key en-
tity, together with its own share. Because of this,
the acting key entity receives all shares of all bits at
the same time. This allows it to reconstruct the orig-
inal bits, without knowing which bit originally be-
longed to which entity, thereby preserving the privacy
of each bit. While this allows the privacy protection
of individual bits, it also leads to an increase in com-
munication complexity. Therefore, using a privacy-
preserving version of this protocol should be justified.
5 SINGLE BROADCAST-BASED
JRNG PROTOCOL
In this section, we present a protocol, which is in-
spired by (Kursawe et al., 2011), that does not require
a secure network and only needs one broadcast per
entity to jointly establish random numbers with a pub-
licly known sum of N. We again assume a group of
at least three entities. Additionally, we assume a pub-
licly known ordering exists. Similar to the protocol
in (Kursawe et al., 2011), this protocol is based on
Diffie-Hellman key exchange and uses the sign func-
tion (Kursawe et al., 2011). However, different from
the protocol by Kursawe et al. our protocol performs
additive aggregation rather than multiplicative aggre-
gation. Thus, we provide the additive variant here.
Before we formalize the protocol, it is important
to note that in this work we assume a fully connected
network. This is, however, not required since all mes-
sages can be publicly known. Therefore, different
network topologies, such as a star network where all
messages are relayed through a semi-honest entity,
can also be used. Additionally, in this work, we set
N = 0, while in practice any N N can be used.
1. A public generator g Z
p
is agreed upon. All
entities generate a random k
p
i
Z
p
and g
k
p
i
.
2. All entities publicly broadcast g
k
p
i
.
3. Each entity p
i
computes its final random number:
R
p
i
=
K
j=1, j6=i
s
i
( j)(g
k
p
j
)
k
p
i
=
K
j=1, j6=i
s
i
( j)g
k
p
i
k
p
j
.
Correctness is easily verified since s
i
( j) + s
j
(i) = 0
for all i 6= j, and thus
K
i=1
R
p
i
=
K
i=1
K
j=1, j6=i
s
i
( j)g
k
p
i
k
p
j
=
K
i=1
i1
j=1
(s
i
( j) + s
j
(i))g
k
p
i
k
p
j
= 0 .
5.1 Security
The security of this protocol is based on the hardness
of Decisional Diffie-Hellman Assumption. As the
protocol is a direct application of the Diffie-Hellman
Key exchange protocol, its security can be reduced to
Decisional Diffie-Hellman problem as R
p
s are con-
structed from the public-key of Diffie-Hellman Ex-
change protocol and can not be distinguished from
random R
p
s.
SECRYPT 2021 - 18th International Conference on Security and Cryptography
440
5.2 Adapted Broadcast-based Protocol
To reduce the complexity of the broadcast-based
scheme, a ring-topology can be assumed. This re-
quires the assumption that any entity p
i
is only able
to communicate with its neighbours: the entities
ranked above and below it. Since this rank is seen in
(mod K), the lowest and highest ranked entities are
also neighbors. The scheme is formalized as follows:
1. A public generator g Z
p
is agreed upon. All
entities generate a random k
p
i
Z
p
and g
k
p
i
.
2. Each entity publicly sends g
k
p
i
to both neighbors.
3. Each entity p
i
computes
R
p
i
=
jN
i
s
i
( j)(g
k
p
j
)
k
p
i
=
jN
i
s
i
( j)g
k
p
i
k
p
j
.
Correctness follows since i N
j
implies that j N
i
.
Therefore, correctness of this scheme can be verified
in the same way as the original scheme is verified.
6 COMPLEXITY ANALYSIS
In this section, we analyze the complexity of our pro-
tocols. For computation complexity, we base our
analysis on the number of modular additions and
modular exponentiation performed. For communica-
tion complexity, we base our analysis on the number
of messages exchanged and their lengths. We con-
clude this section by comparing our protocols.
6.1 Computation Complexity
The share-based protocol is asymmetric since the
number of operations performed by key entities dif-
fers from the number of operations performed by non-
key entities. This underlines the protocol’s versatility:
non-key entities require significantly less computing
power than key entities. Its computation complex-
ity is dominated by modular addition, performed each
round by the acting key entity using all K 1 received
values. For m iterations, this results in m · (K 1)
modular additions. It is important to note that in this
protocol modular addition is performed using single
bits. In contrast to the share-based protocol, the single
broadcast-based protocol is symmetric, meaning that
the number of operations performed is the same for
all entities. Besides modular addition, the complexity
of the broadcast-based protocol is dominated by mod-
ular exponentiation. Each entity performs both oper-
ations once during the protocol: modular exponentia-
tion during the first step and modular addition during
the final step. This results in K modular exponentia-
tions and K ·(K 1) modular additions, since modular
addition is performed using (K 1) numbers. In con-
trast to the first protocol, the addition is performed
using complete numbers. Compared to the original
version of the broadcast-based protocol, its adapted
version has lower computation complexity. This is
caused by the fact that entities only send messages to
their two neighbors. As a consequence, the amount
of numbers used in modular addition is reduced to 2
and computation complexity to 2K. The number of
modular exponentiation remains the same.
6.2 Communication Complexity
For the share-based JRNG protocol, communication
complexity strongly depends on the number of iter-
ations m performed during one run of the protocol.
For each iteration, each entity sends one message to
the acting key entity. Consequently, for m iterations,
this results in m messages send by non-key entities
and m 1 messages send by key entities. For K
entities, this yields a communication complexity of
O(mK). This is comparable to any other system us-
ing the notion of l leaders, as discussed in Related
Work. However, since our protocol requires entities to
share one bit of their random number only, communi-
cation complexity is reduced, even if m far exceeds l.
The single broadcast-based protocol uses broadcast-
ing to communicate. Therefore, its communication
complexity is rather high. Each entity sends one mes-
sage to all others, resulting in K · (K 1) messages
being sent during one execution of the protocol, sim-
plified as K
2
. Additionally, compared to the share-
based protocol, messages are longer, since numbers
are communicated entirely. In the adapted version of
the single broadcast-based protocol, each entity only
sends two messages, thereby reducing communica-
tion complexity to 2K. Similar to the original version
of the protocol, numbers are communicated entirely.
6.3 Comparison
Computation and communication complexity are
shortly summarized in Table 2. When looking at
communication complexity, it becomes visible that
both the share-based and the adapted protocol require
a number of messages that is linear with K. The
share-based protocol requires the shortest messages
and therefore has the lowest communication complex-
ity. When taking into account computation complex-
ity, it becomes visible that, again, the share-based and
adapted protocol have the lowest complexity. Since
the share-based protocol does not require modular ex-
Efficient Joint Random Number Generation for Secure Multi-party Computation
441
Table 2: Computation and communication complexity of existing and our works.
Modular
Addition
Modular
Exp.
Number of
Messages
Message
length (bits)
Encryption/
Decryption
Key
Agreement
(Erkin and Tsudik, 2012) K · (K 1) - K
2
m - -
(Kursawe et al., 2011)
DH key exchange based K · (K 1) 2K K
2
m - -
(Kursawe et al., 2011)
Interactive P · (K 1) - KP m KP/KP -
(Van De Kamp et al., 2019) K · (K 1) - K
2
m - -
(Bell et al., 2020) K · (K 1) - 2K m - 2K
(Hayashi and Koshiba, 2018) K - 2K m - 2K
Our Share-based pro. m · (K 1) - K 1 - -
Our Broadcast-based pro. K · (K 1) K K
2
m - -
Our Adapted broadcast-based pro. 2K K 2K m - -
ponentiation, while the others do, it has the best com-
putation complexity. However, it is important to note
that it does require the existence of secure commu-
nication channels, while messages in the broadcast-
based protocol can be shared publicly.
In Table 2, we also include computation and com-
munication complexity of related works. Comparing
these to the share-based protocol shows that the share-
based protocol has better computation and commu-
nication complexity, as it requires less modular ad-
ditions and sends fewer and shorter messages. Sim-
ilar to (Erkin and Tsudik, 2012), it requires secure
communication channels. Table 2 also shows the
computation and communication complexity of the
broadcast-based protocol are similar to those of the
works by (Erkin and Tsudik, 2012; Kursawe et al.,
2011) and (Van De Kamp et al., 2019). Additionally,
while the works by (Bell et al., 2020) and (Hayashi
and Koshiba, 2018) have better communication com-
plexity, they require additional key agreements to be
performed. The same holds for the interactive proto-
col by (Kursawe et al., 2011), while computation and
communication complexity are similar, it requires ad-
ditional encryption and decryption operations.
7 CONCLUSION
While Joint Random Number Generation can be
found in several Secure Multi-Party Computation so-
lutions, such as data aggregation, most research omits
how to obtain such numbers and simply assumes
their existence. In this paper, we have presented two
computationally efficient protocols for Joint Random
Number Generation. One of these is share-based,
using bit-wise updates to ensure that all individual
numbers yield the correct sum. While this proto-
col does not provide semi-honest security, situations
exist in which the current security guarantees suf-
fice and where computational speed is more impor-
tant. For completeness, we also demonstrate mea-
sures to increase the security of the share-based pro-
tocol. The other protocol is broadcast-based and uses
the sign function to ensure that all generated numbers
yield a public sum. This protocol relies on the com-
putational Diffie-Hellman Problem to provide semi-
honest security. Additionally, we have proposed an
adapted version of the broadcast-based protocol that
has lower computation and communication complex-
ity. By comparing all three protocols based on com-
putation and communication complexity, it becomes
clear that the share-based protocol is most efficient.
However, the fact that it requires secure communi-
cation channels hinders its applicability. Therefore,
when choosing which protocol to employ, the desired
security guarantees and the additional cost of estab-
lishing secure channels should be taken into account.
ACKNOWLEDGMENT
This work was partly supported by SECREDAS
project that has received funding from the Electronic
Component Systems for European Leadership Joint
Undertaking under grant agreement No 783119. This
Joint Undertaking receives support from the European
Union’s Horizon 2020 research and innovation pro-
gramme and Netherlands, Austria, Belgium, Czech
Republic, Germany, Spain, Finland, France, Hungary,
Italy, Poland, Portugal, Romania, Sweden, Tunisia,
United Kingdom.
SECRYPT 2021 - 18th International Conference on Security and Cryptography
442
REFERENCES
Bell, J. H., Bonawitz, K. A., Gasc
´
on, A., Lepoint, T., and
Raykova, M. (2020). Secure single-server aggregation
with (poly) logarithmic overhead. In Proceedings of
the 2020 ACM SIGSAC Conference on Computer and
Communications Security, pages 1253–1269.
Erkin, Z. and Tsudik, G. (2012). Private computation of
spatial and temporal power consumption with smart
meters. In International Conference on Applied
Cryptography and Network Security, pages 561–577.
Springer.
Erkin, Z., Veugen, T., Toft, T., and Lagendijk, R. L. (2013).
Privacy-preserving distributed clustering. EURASIP
Journal on Information Security, 2013(1):4.
European Parliament and Council of European Union
(2016). Regulation (eu) 2016/679.
Availableat:https://eur-lex.europa.eu/legal-content/
EN/TXT/HTML/?uri=CELEX:32016R0679&from=
EN.
Garay, J., Schoenmakers, B., and Villegas, J. (2007). Prac-
tical and secure solutions for integer comparison. In
International Workshop on Public Key Cryptography,
pages 330–342. Springer.
Garcia, F. D. and Jacobs, B. (2010). Privacy-friendly
energy-metering via homomorphic encryption. In In-
ternational Workshop on Security and Trust Manage-
ment, pages 226–238. Springer.
Hayashi, M. (2018). Secure modulo sum via multiple access
channel. arXiv preprint arXiv:1812.10862.
Hayashi, M. and Koshiba, T. (2018). Secure modulo zero-
sum randomness as cryptographic resource.
Hayashi, M. and Koshiba, T. (2019). Verifiable quan-
tum secure modulo summation. arXiv preprint
arXiv:1910.05976.
Jiang, L., Lou, X., Tan, R., and Zhao, J. (2019). Differen-
tially private collaborative learning for the iot edge. In
EWSN, pages 341–346.
Kursawe, K., Danezis, G., and Kohlweiss, M. (2011).
Privacy-friendly aggregation for the smart-grid. In In-
ternational Symposium on Privacy Enhancing Tech-
nologies Symposium, pages 175–191. Springer.
Lu, R., Heung, K., Lashkari, A. H., and Ghorbani, A. A.
(2017). A lightweight privacy-preserving data aggre-
gation scheme for fog computing-enhanced iot. IEEE
Access, 5:3302–3312.
Lyu, L. (2020). Lightweight crypto-assisted distributed
differential privacy for privacy-preserving distributed
learning. In 2020 International Joint Conference on
Neural Networks (IJCNN), pages 1–8. IEEE.
Merkle, R. C. (1978). Secure communications over insecure
channels. Communications of the ACM, 21(4):294–
299.
Privacy International (2020). Data breach.
Availableat:https://privacyinternational.org/examples/
data-breach.
Shi, E., Chan, T. H., Rieffel, E., Chow, R., and Song, D.
(2011). Privacy-preserving aggregation of time-series
data. In Proc. NDSS, volume 2, pages 1–17. Citeseer.
Shi, Z., Sun, R., Lu, R., Chen, L., Chen, J., and Shen, X. S.
(2015). Diverse grouping-based aggregation protocol
with error detection for smart grid communications.
IEEE Transactions on Smart Grid, 6(6):2856–2868.
Singapore Parliament (2014). Personal data protection act
overview.
Availableat:pdpc.gov.sg/Overview-of-PDPA/
The-Legislation/Personal-Data-Protection-Act.
Van De Kamp, L., Ugwuoke, C., and Erkin, Z. (2019).
Economy: Ensemble collaborative learning using
masking. In 2019 IEEE 30th International Sympo-
sium on Personal, Indoor and Mobile Radio Commu-
nications (PIMRC Workshops), pages 1–6. IEEE.
Wang, W. and Lu, Z. (2013). Cyber security in the smart
grid: Survey and challenges. Computer networks,
57(5):1344–1371.
Yao, A. C. (1982). Protocols for secure computations. In
23rd annual symposium on foundations of computer
science (sfcs 1982), pages 160–164. IEEE.
Zhao, L., Wang, Q., Zou, Q., Zhang, Y., and Chen, Y.
(2019). Privacy-preserving collaborative deep learn-
ing with unreliable participants. IEEE Transactions on
Information Forensics and Security, 15:1486–1500.
Efficient Joint Random Number Generation for Secure Multi-party Computation
443