Privacy Preserving Reputation Calculation in P2P Systems with Homomorphic Encryption

International Journal of Computer Networks & Communications (IJCNC) Vol.13, No.6, 2021
Privacy Preserving Reputation Calculation in
P2P Systems with Homomorphic Encryption
FUJITA Satoshi
Graduate School of Advanced Science and Engineering, Hiroshima
University, Kagamiyama 1-4-1, Higashi-Hiroshima, 739-8527, Japan
Abstract. In this paper, we consider the problem of calculating the node reputation in a Peer-to-
Peer (P2P) system from fragments of partial knowledge concerned with the trustfulness of nodes
which are subjectively given by each node (i.e., evaluator) participating in the system. We are
particularly interested in the distributed processing of the calculation of reputation scores while
preserving the privacy of evaluators. The basic idea of the proposed method is to extend the
EigenTrust reputation management system with the notion of homomorphic cryptosystem. More
specifically, it calculates the main eigenvector of a linear system which models the trustfulness
of the users (nodes) in the P2P system in a distributed manner, in such a way that: 1) it blocks
accesses to the trust value by the nodes to have the secret key used for the decryption, 2) it improves
the efficiency of calculation by offloading a part of the task to the participating nodes, and 3) it
uses different public keys during the calculation to improve the robustness against the leave of
nodes. The performance of the proposed method is evaluated through numerical calculations.
Keywords: P2P reputation management, homomorphic cryptosystem, EigenTrust, Paillier cryp-
tosystem.
1 Introduction
Reputation management is a crucial task for fully distributed systems such as Peer-
to-Peer (P2P) systems and federated social networks since in those systems, services
are provided by individual, unauthorized peers (i.e., nodes) unlike classical server-
based systems. The reputation of nodes is used not only to find high quality services
as in gourmet sites, but also to identify low quality service-providers to penalize
them by deliberately reducing the service quality supplied to them. Indeed, many
P2P systems support reputation management as a part of the incentive mechanism,
so as to encourage nodes to increase their reputation by improving the service
quality supplied by those nodes.
The reputation of a node in such systems is generally calculated from trust
value of nodes given by entities called evaluators (e.g., customers in the case of
online shopping site). Therefore, the trust value can be an obvious target of attacks
by malicious nodes who wish to Illegally raise their reputation. One way to protect
such attacks is to ask trustworthy third parties to calculate trust value of the
evaluatee and to keep it securely, where if trust values calculated by different entities
DOI: 10.5121/ijcnc.2021.13601 1

conflict, an agreement could be achieved through majority voting. Such a consign-
based approach would work well if the trust value of a node can be (automatically)
generated in an objective manner, as in cases in which the contribution of an amount
of upload bandwidth in parallel download systems such as BitTorrent and the
forwarding of received video streams in video streaming systems such as seedess1
and P2PSP2 (if the credibility of the prepared server is questionable, we could
record the history of contributions and reputation calculations as a blockchain, so
that the recorded information can be verified by all nodes involved in the service).
In this paper, we consider a more general setting for the reputation management
in P2P systems in which the trust value of an evaluatee is given by an evaluator
in a subjective manner and the evaluator wishes to keep it secret from other nodes
including the evaluatee (e.g., the reader could consider the personnel evaluation
in an organization such as company and alumni meeting). To the author’s best
knowledge, existing reputation management schemes for P2P systems do not take
such situations into account at all and do not adequately preserve the privacy of the
evaluators. In this paper, we tackle this challenging issue, and propose a method
to calculate the reputation of each node in a distributed manner without disclosing
any subjective trust value, with the aid of homomorphic encryption.
There are a lot of previous work concerned with the privacy preserving cal-
culations with homomorphic encryption [11–13, 16]. Most of them assume that a
trustworthy third party is commissioned to conduct the calculation on confidential
data stored in a cloud server, which is encrypted with the public key of an appro-
priate key server. Thus, it implicitly allows the trustworthy key server to decrypt
the data even if it was strictly confidential.
Such a centralized approach, however, does not work well in P2P systems for
the following reasons. At first, such calculations should be realized so as to block
accesses to the trust values by a node which has the secret key for the decryption.
Specifically, although it could access the ciphertext of the reputation score which
is the final result of calculation, it should not be allowed to access any data from
which the original trust value of nodes could be inferred. Concerned with this issue,
in the proposed method, we take an approach so that encrypted reputation values
are successively updated through calculation to reflect the trust value of individual
nodes. Although the intermediate results of the calculation are shared by all nodes,
the (original) trust values cannot be guessed from them besides the fact that the
sum of trust values given by an evaluator equals to a certain fixed value. The sec-
ond point we need to consider is the decentralization of the reputation calculation.
Existing privacy preserving schemes are designed for the bulk processing in which
a central entity conducts the calculation on the data stored on a cloud server. In
contrast, the data on P2P systems, including trust values, are distributed over the
1
https://seedess.com/
2
https://p2psp.org/
2

network from the beginning, and it is not efficient to aggregate them to a central-
ized server before starting the calculation. In addition, the calculation time can
significantly be reduced by utilizing the computational resources of the participat-
ing nodes. In this aspect, the proposed method incorporates several techniques to
improve the efficiency of the distributed processing. The third point is that due to
node churn, it is not guaranteed that a node holding the secret key will stay in
the system at the time of decryption. To overcome this issue, we use different pub-
lic keys for the reputation calculation. The effectiveness of the proposed method is
evaluated by conducting numerical calculations. The results show that the proposed
method reduces the maximum circulation time used for aggregating the result of
multiplications to a half, thereby reducing the time required for each round of the
reputation computation, and the cost of privacy preservation is proportional to the
number of nodes N, which takes about 16 seconds when N = 1024.
The remainder of this paper is organized as follows. Section 2 overviews related
work. Section 3 reviews the EigenTrust reputation management system which plays
a central role in the proposed method. Section 4 describes the details of the pro-
posed method. Section 5 summarizes the results of evaluations. Finally, Section 6
concludes the paper with future work.
2 Related Work
2.1 Homomorphic Encryption
Homomorphic encryption (HE, for short) is a type of encryption which preserves
the homomorphism for certain arithmetic operations3 and is classified into several
types by the strength of preservations, such as partially homomorphic, somewhat
homomorphic, leveled fully homomorphic, and fully homomorphic. The concept of
HE was initially proposed by Rivest in 1978 [14]. Although the concrete realization
of HE has been an open issue for thirty years, it was positively answered by Craig
Gentry in 2009 [5] with the proposal of a fully homomorphic cryptosystem based
on the lattice-based cryptography.
As a tool available to developpers who are not a specialist in cryptography,
IBM released a C++ library named HElib[6]4, which is being updated actively.
The current version of HElib adopts the Brakerski-Gentry-Baikuntanathan (BGV)
method and the Cheon-Kim-Kim-Song (CKKS) method, which are extentions of
the Gentry’s method. The number of options available for the implementation of
security systems dramatically increases if we do not stick to be fully homomorphic.
3
Homomorphism in the context of HE implies the property such that the decryption of the
result of an operation applied to ciphertexts is consistent with the result applied to the original
plaintexts.
4
https://github.com/homenc/HElib
3

Table 1. Execution time required for the encoding/decoding in the Paillier cryptosystem. Two
values in each cell show the mean and the sample standard deviation over 1000 runs, respectively.
Bit length Key generation [ms] Encoding [ms] Decoding [ms]
1024 19.32 7.72 (0.67) 7.64 (0.69)
2048 27.41 7.65 (0.59) 7.57 (0.63)
4096 641.58 47.54 (3.62) 47.29 (3.72)
8192 12782.64 273.42 (16.83) 272.87 (16.50)
In fact, it is widely known that RSA and ElGamal5 are multiplicative HE, and
Goldwasser-Micali and Paillier [10] are additive HE, where the last one is the main
player of the current paper. A typical application for additive HEs is the electronic
voting, as it merely needs simple counting of votes. Other applications of additive
HEs include the analysis of medical data [1] and the k-clustering of vector data
[18]. A formal definition of the Paillier cryptosystem is given in Appendix to make
this paper self contained, and before proceeding to the detailed explanation of the
proposed method, as a preliminary experiment, we evaluated the execution time
of a Python program written with the paillierlib library on a computer with Intel
Core i9, 2.3 GHz, and 16 GB memory. Table 1 summarizes the results, where the
key length is varied from 1024 bits to 8192 bits. From the table, we find that when
the key length is 2048 bits, encoding and decoding operations can be done in 10
ms each, confirming that the Paillier cryptosystem can be used as a tool for the
privacy preserving calculations.
2.2 Related Work on Reputation Management
In the literature, there are many proposals concerned with the reputation manage-
ment in P2P systems. PeerTrust [15] is designed for the transaction-based feedback
system and calculates the reputation of each node by considering three basic trust
parameters and two adaptive factors, which include the amount of feedback received
from other nodes, the number of transactions relevant to the node, the credibility
of the source of feedback, and the context of transactions and the community.
PowerTrust [17] takes advantage of the distribution of evaluations among nodes,
which is generally not uniform but follows a power-law. Specifically, it adaptively
selects a small number of nodes with the highest reputation, and leverages them to
significantly improve the accuracy of calculation and the speed of convergence.
Although a detailed explanation of EigenTrust will be given in the next section,
we could overview several proposals relevant to EigenTrust, as follows. Choi et al. [3]
proposed Enhanced Eigentrust, which uses a beta distribution to calculate the nor-
malized trust values. The intensive employment of trusted nodes effectively speed
up the convergence in calculating reputation values. Personalized EigenTrust [2]
5
ElGamal cryptosystem is a public-key cryptosystem whose difficulty is based on the discrete
logarithm problem.
4

allows each node to individually specify its trusted nodes and the method proposed
in [4] recognizes such trusted nodes with the aid of Page Rank. Federated Eigen-
Trust [7] introduces the notion of representative nodes to manage the unpredictable
leave of trusted nodes. HonestPeer [9] oﬄoads the task of reputation calculation to
(honest) nodes with high ratings to legitimately increase the reputation of nodes.
3 Reputation Calculation with EigenTrust
EigenTrust provides a method to calculate the (global) reputation value of each
node in a P2P system based on the (local) trust value given by each node to other
nodes. Note that such a collection of trust values can be represented in the form of
a square matrix. In the following, we assume that the trust value cij given by node
i for j is normalized to the [0, 1] interval so that the higher the value, the higher
the trust level6. It is natural to assume that the value of cjk is trustable for i, if
cij is suﬃciently large (each node is assumed to have absolute trust in itself and
cii = 1 holds for any i). In other words, we could assume the existence of a sort of
transitivity concerned with the trust. By extending this idea, given a collection of
local trust values, the reputation tik of i in k could be represented as
tik :=
n
∑
j=1
cijcjk.
In other words, we have





ti1
ti2
.
.
.
tin





:=





c11 c21 · · · cn1
c12 c22 · · · cn2
.
.
.
.
.
.
.
.
.
c1n c2n · · · cnn










ci1
ci2
.
.
.
cin





= (⃗
c1,⃗
c2, · · · ,⃗
cn)⃗
ci = C⃗
ci,
where ⃗
ci = (ci1, . . . , cin)T . Although the above ⃗
ti = (ti1, ti2, . . . , tin)T is the repu-
tation of nodes through a chain of length two from i, it can be easily extended to
the chain of length n so that ⃗
ti := Cn⃗
ci. Vector ⃗
ti is known to converge to the main
eigenvector of matrix C as n → ∞ (i.e., to the eigenvector such that the eigenvalue
is one) regardless of the value of ci, as long as matrix C is irreducible and acyclic,
and the existence of such an eigenvector is guaranteed by the Peron Frobenius
Theorem. It is worth noting that ⃗
t = limn→∞ Cn⃗
c corresponds to the stationary
distribution of Markov chain in such a way that the transition probability between
states is given by C. Although such ⃗
t can be obtained algebraically by solving the
6
In this setting, we could not distinguish between not trusting j and not having an opinion about
j.
5

eigenvalue equation C⃗
t = ⃗
t, since the computational complexity increases as n in-
creases, it is common to solve this equation numerically by repeatedly multiplying
C to an appropriate initial vector.
4 Proposed Method
There are several options to apply the notion of homomorphic cryptosystem to
the reputation management in EigenTrust. One possible idea is to encrypt trust
value cij and to multiply it with ei after collecting encrypted values from all nodes.
However, such a naive approach does not scale since it forces the central entity to
conduct all multiplications, and is less efficient than a simple centralized approach
in which all encrypted cij’s are collected to the server before starting the calculation.
Thus in the following, we will focus our attention on another approach in which
the encryption is applied to reputation vector ⃗
e.
4.1 Baseline
At first, we consider a simple client-server (C/S) scheme, which will be used as
the baseline in the proposed method. In this scheme, a central server manages
the encrypted vector ⃗
e = (e1, e2, . . . , en) and repeats the following steps until ⃗
e
converges to a certain reputation vector:
1. For each i, the server sends ei to client i;
2. Client i calculates ηj := ei × cij for all j, and replies the result to the server;
and
3. After receiving encrypted values from all clients, the server updates ⃗
e as ei :=
∑
j ηj for each i.
Note that since the fact of cij = 0 is hidden by applying the encryption, exactly n2
multiplications should be conducted even if the given matrix is sparse, unless the
fact of cij = 0 is notified to the server through alternative route.
Since the role of client i is to receive encrypted ei from the server and to return
the resulting list (ei×ci1, . . . , ei×cin) to the server, if the multiplication is conducted
under the Paillier cryptography, merely the results of multiplications to the trust
values are collected to the server without revealing its own trust values. In addition,
since an updated ei will be received from the server in the next round, two successive
rounds are separated through the barrier synchronization.
4.2 Distributed Aggregation
The above C/S scheme can be extended so that the aggregation of the results is
replaced by the P2P communication. More concretely, we organize the task of ag-
gregating the result of n × n independent multiplications into several task-groups
6

1 2 3 4 5 6
7 8 9 10 11 12
13 14 15 16 17 18
19 20 21 22 23 24
U1
U2
U3
U4
V1 V2 V3 V4 V5 V6
Fig. 1. Two partitions of the set of nodes V = {1, 2, . . . , 24} with k = 4. Four subsets in U
are represented by dashed blue rectangles and six subsets in V are represented by dashed red
rectangles.
and to give the responsibility of controlling each task-group to a node in V in a
decentralized manner. Note that such an extension of data aggregation still pre-
serves an important nature of the baseline scheme so that each node does not have
to disclose its trust values to the others.
Assume n(= |V |) is not a prime and let k be a divisor of n. At ﬁrst, we partition
V into n/k subsets of an equal size as V = {Vi : 1 ≤ i ≤ n/k}, and at the same
time, partition V into k subsets of an equal size as U = {Uj : 1 ≤ j ≤ k} (the reader
could imagine two orthogonal partitions naturally realized from a two-dimensional
array of size k × n/k. See Figure 1 for illustration). Let Mi = {ei × cij : 1 ≤ j ≤ n}
be the set of n independent multiplications conducted by node i in the baseline
scheme. We consider a partition of Mi into n/k subsets Mi,1, . . . , Mi,n/k, so that
Mi,y consists of multiplications for nodes j in subset Vy. See Figure 2 for illustration.
Then for each Ux ∈ U, we can organize a collection of task groups in such a way
that the jth task-group consists of the aggregation of the result of multiplications
in
∪
i∈Ux
Mi,j. Since subset Ux consists of n/k nodes, we can assign such n/k task-
groups to the nodes in Ux in one-to-one manner. Let Tx,y be the task-group assigned
to the yth node in subset Ux. Note that the outcome of Tx,y is a collection of results
of k independent inner products of sub-vectors of length n/k each, which are all
conducted by the nodes in subset Ux. Although the total number of multiplications
controlled by a node in Ux does not change even if the value of parameter k is
varied, a too small k increases the number of additions to be conducted at the
centralized server. See Figure 3 for illustration.
The concrete execution of task-group Tx,y proceeds as follows. Suppose that
each node i computes multiplication ei × cij by using encrypted ei received from
7

1 2 3 4 5 6
7 8 9 10 11 12
13 14 15 16 17 18
19 20 21 22 23 24
U1
U2
U3
U4
M9,1 M9,2 M9,3 M9,4 M9,5 M9,6
Fig. 2. 24 multiplications concerned with node 9 in subset U2 ∈ U. Nodes in U2 are represented
by circles and multiplications conducted by node 9 are represented by simple numbers. Set of
multiplications M9,y, for 1 ≤ y ≤ 6, is indicated by a dashed red rectangle which corresponds to
task group T2,y since node 9 is a member of subset U2.
the central entity, and stores the results locally. The aggregation of those results
is independently done for each j contained in Tx,y by using a chain of length at
most n/k (a part of such chain is indicated by a blue arrow in Figure 3). The
order of nodes in the chain is determined by the responsible node, say i∗, and the
aggregation is realized by flowing a message along the determined route as follows:
At first, the responsible node i∗ sends out ηi∗ := ei∗ × ci∗j to the first node on the
chain. Upon receiving an encrypted value from the predecessor, node i′ adds its
own ei′ × ci′j to the received value, and passes the result to the next node on the
chain, where the last node on the chain returns the result to i∗. The responsible
node i∗ can initiate such a flow on several independent chains if n/k is large, while
in such a case, to prevent the double counting of ηi∗ , node i∗ should send out ηi∗
only for one chain and value 0 for the remaining chains.
4.3 Adaptive Load Balancing
In the above scheme, several messages are circulated along chains in each subset
Ux ∈ U to conduct the partial update of tentative reputation value of nodes in V .
The final reputation vector is obtained by repeating rounds similar to EigenTrust,
and the next round can start only after collecting partial updates from all subsets.
Thus we could speed up the reputation calculation by reducing the maximum cir-
culation time over all chains while keeping the number chains (recall that the C/S
scheme described before trivially achieves the min-max circulation time). In the
proposed method, we realize such a reduction of the maximum circulation time by
using the following two techniques:
8

22
22
22
22
22
22
M7
M12
M11
M10
M9
M8
Fig. 3. Aggregation of the result of multiplications through circulating an aggregation massage.
In this figure, each layer corresponds to n multiplications conducted by a node in U2, where the
results for node 22 are placed at the same location in each layer. Thus, the result of partial inner
product of length n/k can be obtained by aggregating the result of n/k multiplications conducted
by six nodes in U2.
– Organize U so that each subset Ux in the partition consists of nodes whose
mutual distance is short.
– The responsible node i∗ monitors the circulation time of each chain managed
by i∗, and re-organize the collection of chains so that the maximum circulation
time could be minimized.
A typical realization of the first technique is to measure the round-trip time for
each pair of nodes in V beforehand, and to apply the k-means method to obtain
a collection of k subsets. On the other hand, the second technique can be realized
by adaptively applying split and merge operations to the collection of chains in
such a way that: 1) a chain with long circulation time is split into two chains and
2) two chains of small circulation time are merged into a chain. Note that if the
partition U obtained by the first technique reflects the proximity of the nodes, we
could assume that the transmission time to the next node on any chain derived
from U is bounded by a sufficiently small value such as 50ms, and given such a
subset of nodes, the second technique manages the difference of response time due
to the overload of nodes.
4.4 Improve the Resilience Against Churn
In the previous description, we assumed that each element of the reputation vector
is encrypted using the public key of the central key server, but this assumption can
be relaxed as follows:
– Different keys can be used for each round, if it is allowed to decrypt the repu-
tation vector after each round.
9

Central server
Responsible node A Responsible node B
Task group Tx,y Task group Tx’,y’
Request for decryption
Ciphertext of ej encrypted
with public key of B
Plaintext of partially
updated ej
Reflecting the local trust
values for node j to the
accumulated value
Fig. 4. The flow of reputation calculation in a round. Although only two task groups are illustrated
in this figure, there are n(= k ×n/k) task-groups in the system each of which aggregates the result
of n multiplications. Note that the central server is used to collect the resulting reputation scores
and the actual calculation is offloaded to the participants.
– Different keys can be used for each task group.
– The key used to decrypt the outcome of a task group does not have to be held
by the responsible node of the task group. In fact, the task of decryption can
be delegated to the responsible node of other task group.
With such a refinement, the description of the overall task group is modified as
follows (see Figure 4 for illustration):
1. The responsible node i∗ of a task group receives the tentative reputation value
of each member from the server, and encrypts them using the public key of
the node which is responsible for the decryption, or the server conducts the
encryption before sending it to node i∗.
2. The responsible node i∗ sends the encrypted tentative reputation value to each
member in the subset, conducts the partial aggregation under its control, and
forwards the resulting list to the node responsible for decryption, where the
resulting list contains n elements.
3. After receiving a list to be descrypted, the node decrypts each element in the
list, and returns the result to the central server.
In the above extension, the grouping of nodes causes no load balancing effect at all
but increases the total load of the nodes since each responsible node must decrypt a
10

list of length n every round. Instead, the content of calculations in each task group
is kept secret from the server and members of other subsets, which increases the
level of privacy preserving compared to the C/S scheme. In addition, by delegating
the task of decryption to the members of other subset, we could preserve the privacy
of each member from other members in the same group including the responsible
node i∗. Finally, since the role of key holders is only to decrypt the result of a task
group in a certain round, it can be easily replaced by other node in each round,
thereby increasing the resilience against node churn.
5 Evaluation
In this section, we evaluate the performance of the proposed method, in terms
of: 1) the effect of acceleration of the data aggregation and 2) the amount of ad-
ditional cost due to privacy preservation in P2P environment through numerical
calculations. More specifically, as for the first point,
– We evaluate how much the circulation of aggregation message can be accelerated
by using the k-means clustering and a heuristic minimization of the length of
the cyclic route, by assuming that nodes are associated with a point in the
three-dimensional Euclidean space, and
– We evaluate how much the maximum circulation time over all clusters can be
reduced by reflecting the variance of the response time of nodes, by assuming
that the response time follows a geometric distribution.
The reader should note that the node density around a node becomes sparse as
the number of dimensions increases, which contradicts to the intuition such that
the locality of nodes plays a crucial role in organizing efficient P2P overlay, and
if the dimension is fixed to two, it becomes difficult to reflect the diversity of the
locality existing in real-world networks. On the other hand, the slow-down due to
privacy preservation will be estimated in Section 5.2 using the result of preliminary
experiments summarized in Table 1.
5.1 Effect of Reflecting the Locality of Nodes
At first, we evaluate the effect of the k-means clustering to reduce the maximum cir-
culation time (over all subsets) by assuming that each responsible node determines
the cyclic route in a random manner. Figure 1 summarizes the results. The vertical
axis indicates the ratio of the maximum circulation time obtained by the k-means
method to the value obtained by applying a random partitioning (into subsets with
an equal number of nodes), and the height of each bar represents the value averaged
over 100 random instances generated for each combination of the number of nodes
N and the number of clusters k. We confirm that no instance yields a ratio worse
than 1.0, and if k ≥ 8, the k-means clustering reduces the maximum circulation
11

512 1024 2048 4096 8192
Number of Nodes
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Ratio
of
Maximum
Circulation
Time
Efect of k-Means Clustering Compared with Random Partitioning
k=4
k=8
k=16
k=32
k=64
k=128
Fig. 5. Speed-up of the aggregation time by the k-means clustering compared with a random
clustering into subsets of an equal size.
time of the random partitioning to almost a half, on average. Such a significant
reduction should be caused by the short average distance in the resulting clusters,
since the cluster size obtained by the k-means method is not necessarily equal (i.e.,
the maximum cluster size is larger than the random partitioning). To clarify this
point, we evaluated the maximum average distance over all subsets generated by
two schemes. Figure 6 summarizes the results. The average distance realized by
the random partitioning gradually increases as k increases, which is because we are
taking the maximum circulation time over k subsets. In the k-means method, on
the other hand, since the effect of localization increases for larger k’s, the ratio to
the random partitioning becomes smaller than the ratio indicated in Figure 5 (e.g.,
the ratio is less than 0.25 for k = 256).
Next, we evaluate the effect of circulation order to the aggregation time, which
is optimized in the proposed method by using an approximation scheme based on
the minimum spanning tree (MST). As the result of experiments, we confirmed
that although the length of the cycle, i.e., summation of the edge weights, realized
by a random ordering increases in proportion to the number of nodes in the cluster,
the cycle length realized by the proposed method increased about 1.58 times when
12

4 8 16 32 64 128 256
Cluster size K
0
10
20
30
40
50
60
70
Average
distance
among
nodes
within
a
cluater
[ms]
Maximum average distance over all clusters (N=8192)
k-means
random
Fig. 6. Maximum average distance over all clusters.
the cluster size doubled, on average. This value of 1.58 is roughly in line with the
degree of increase in the weight of the MST for the given instance. In fact, while it
is well known that the theoretical upper bound on the approximation ratio of the
MST-based approximation scheme is 2.0, it was experimentally confirmed that the
length of approximated solution covering randomly placed points in a 3D cube is,
1.2 to 1.3 times the weight of the MST (this ratio is slowly worsened as the number
of nodes increases). In terms of the real-time, the above result roughly corresponds
to the situation in which the aggregation time of 17 seconds taken by the random
ordering reduces to about 4 seconds by the proposed method, when N = 512.
The effect of the load balancing reflecting the variance of the response time of
nodes is evaluated as follows. In this experiment, we consider a theoretical model
in which the response time of each node follows a geometric distribution with prob-
ability p. Then, we evaluated to what extent the maximum sum of the response
times over all subsets can be reduced by using a greedy partitioning scheme instead
of a random partitioning, by varying the number of nodes N and the probability p
(note that the smaller p is, the mean and the variance of the response times become
larger). The results are summarized in Figure 7. We confirm that as N increases,
13

64 128 256 512 1024 2048
Number of Nodes in a Cluster
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
Ratio
of
Accumulated
Response
Time
to
the
Randomized
Scheme Effect of Greedy Load Balancing
p=0.5
p=0.3
p=0.1
p=0.01
Fig. 7. Effect of load balancing by the greedy scheme.
the room for the improvement by the greedy scheme becomes smaller since the bad-
ness of the random partitioning will be amortized. However, even with N = 128,
we can obtain a speed-up of more than 4% for any p ≤ 0.5.
5.2 Cost of Privacy Preservation
Finally, we evaluate the increase of the computation time due to the enhancement
of privacy preservation. In the baseline method, encryption is conducted only once
at the beginning of computation and decryption is conducted only once when the
final reputation value is obtained. In the enhanced method, on the other hand,
encryption and decryption are conducted for all reputation values with different
keys in each round to prevent other nodes from eavesdropping the trust values.
Thus, the additional overhead per round can be estimated as the time required
to encrypt/decrypt one element multiplied by the number of elements. The results
in Table 1 imply that when the number of elements is N, the time required for
encryption and decryption is 7.65N [ms] and 7.57N [ms], respectively; e.g., when
N = 1024, it takes about 8 seconds for encryption and decryption, respectively.
14

Note that this cost can be mitigated by conducting the encryption/decryption once
every few rounds, rather than every round.
6 Concluding Remarks
In this paper, we propose a method to extend the EigenTrust reputation manage-
ment system with the notion of homomorphic cryptosystem so that the privacy
of evaluations is protected from other nodes. Experimental results show that the
proposed method reduces the maximum circulation time used for aggregating the
result of multiplications to a half, thereby reducing the time required for each round
of the reputation computation, and the cost of privacy preservation is proportional
to the number of nodes N, which takes about 16 seconds when N = 1024. A future
work is to implement the proposed method in existing P2P systems.
Conﬂicts of Interest
The author declares no conﬂict of interest.
References
1. A. Ara, M. Al-Rodhaan, Y. Tian, and A. Al-Dhelaan. A secure privacy-preserving data
aggregation scheme based on bilinear ElGamal cryptosystem for remote health monitoring
systems. IEEE Access, 5, pp. 12601–12617, 2017.
2. N. Chiluka, N. Andrade, D. Gkorou, and J. Pouwelse. Personalizing EigenTrust in the Face of
Communities and Centrality Attack. in Proc. IEEE 26th Int’l Conf. on Advanced Information
Networking and Applications, 2012.
3. D. Choi, S. Jin, Y. Lee, and Y. Park. Personalized eigentrust with the beta distribution.
ETRI Journal, 32(2):348–350, 2010.
4. P. A. Chirita, W. Nejdl, M. T. Schlosser, and O. Scurtu. Personalized Reputation Manage-
ment in P2P Networks. in Proc. ISWC Workshop on Trust, Security, and Reputation on the
Semantic Web, 2004.
5. C. Gentry. Fully Homomorphic Encryption Using Ideal Lattices. in Proc. the 41st ACM
Symposium on Theory of Computing (STOC), 2009.
6. S. Halevi and V. Shoup. Algorithms in HElib. In Proc. CRYPTO 2014: Advances in Cryp-
tology, 2014, pages 554–571.
7. R. Jansen, T. Kaminski, F. Korsakov, A. S. Croix, and D. Selifonov. A Priori Trust Vulner-
abilities in EigenTrust. Technical report, University of Minnesota, 2008.
8. S.D. Kamvar, M.T. Schlosser, H. Garcia-Molina. The Eigentrust algorithm for reputation
management in P2P networks. in Proc. 12th Int’l Conf. on World Wide Web (WWW ’03),
2003, pages 640–651.
9. H. A. Kurdi. HonestPeer: An enhanced EigenTrust algorithm for reputation management in
P2P systems. Journal of King Saud University – Computer and Information Sciences (2015)
27, 315–322.
10. P. Paillier. Public-Key Cryptosystems Based on Composite Degree Residuosity Classes. in
Proc. EUROCRYPT 1999, pages 223–238.
11. M. M. Potey, C. A. Dhote, and D. H. Sharma. Homomorphic encryption applied to the cloud
computing security. Procedia Computer Science, 79, pp. 175–181, 2016.
15

12. M. Tebaa, S. E.Hajji, and A. E. Ghazi. Homomorphic encryption method applied to Cloud
Computing in Proc. National Days of Network Security and Systems, 2012.
13. M. Tebaa and S. E. Hajji. Secure cloud computing through homomorphic encryption. Com-
puter Science, ArXiv, Corpus ID: 21074700, 2014.
14. R. L. Rivest, L. Adleman, and M. L. Dertouzos. On data banks and privacy homomorphisms.
In Foundations of Secure Computation, 1978.
15. L. Xiong and L. Liu. PeerTrust: supporting reputation-based trust for node-to-node electronic
communities. IEEE Trans. Knowledge and Data Engineering, 16(7):843–857, 2004.
16. F. Zhao, C. Li, and C. F. Liu. A cloud computing security solution based on fully homomor-
phic encryption. in Proc. 16th Int’l Conf. on Advanced Communication Technology, 2014.
17. R. Zhou and K. Hwang. PowerTrust: A Robust and Scalable Reputation System for Trusted
Peer-to-Peer Computing. IEEE Trans. Parallel and Distributed Systems, 18(4):460–473,
2007.
18. Y. Zhu and X. Li. Privacy-preserving k-means clustering with local synchronization in peer-
to-peer networks. Peer-to-Peer Networking and Applications, 13, pp. 2272–2284 (2020).
A Paillier Cryptosystem
To make this paper self-contained, we give a formal definition of Paillier cryptosys-
tem which plays a crucial role in the proposed method.
A.1 Euler’s Theorem
We begin with an important theorem in the number theory, which is known as the
Euler’s theorem.
Euler’s Theorem For any mutually prime positive integers n and a, it holds
aϕ(n)
≡ 1 (mod n)
where ϕ(n) is the Euler’s totient function which returns the number of integers in
{1, 2, . . . , n} which are mutually prime to n.
As an extension of the Euler’s totient function, Paillier cryptosystem uses a
function λ(n) satisfying aλ(n) ≡ 1 (mod n) for any mutually prime integers n and
a. By the Carmichael’s theorem, function λ(n) is represented as
λ(
∏
i
pei
i ) = lcmi(λ(pei−1
i ))
by using the prime factorization of n =
∏
i pei
i , where lcmi denotes the least common
multiple for all i (note that lcmi(pei
i ) = n holds by definition) and each λ(pe) is
defined so that when p = 2,
λ(2e
) =



1 if e = 1
2 if e = 2
2e−2 if e ≥ 3
16

and for odd prime p,
λ(pe
) = pe−1
(p − 1).
In particular, since n =
∏
i pi implies
λ(n2
) = lcmi(λ(p2
i )) = lcmi(pi × (pi − 1))
=
∏
i
pi × lcmi(pi − 1) = nλ(n),
it holds
aλ(n2)
≡ anλ(n)
≡ 1 (mod n2
) (1)
where the last equation is used to certify the correctness of the decryption in Paillier
cryptosystem.
A.2 Encryption/Decryption
Paillier cryptosystem is a public key cryptosystem. Secret key, which is used for
the decryption, is a pair of sufficiently large prime numbers (p, q), and public
key (g, n), which is used for the encryption, is generated as n := p × q and g :=
(kn + 1) mod n2, where k is a random number drawn from Zn. Given a public key
(g, n), the ciphertext c of a message m satisfying 0 ≤ m < n is calculated as
c := gm
· rn
mod n2
,
where r is a random number in Zn which is independently selected for every time
of the encryption. On the other hand, the decryption of a given ciphertext c is
conducted as:
m :=
L(cλ mod n2)
L(gλ mod n2)
mod n
where L(u) = u−1
n and λ is a number defined as
λ = lcm(p − 1, q − 1).
Recall that by the Carmichael’s theorem, aλ ≡ 1 (mod n) holds for any mutually
prime integers a and n, and since c and g are residuals of a division by n2, a and
n are certainly mutually prime. The decryption of c uses λ which is calculated
from the secret key (p, q), but it is secured by the fact that the factorization of n
is computationally hard. Now let us verify whether the above procedure certainly
decrypts c to plaintext m. At first, cλ is calculated as
cλ
= (gm
· rn
)λ
mod n2
⇐ Def. of c
= gmλ
rnλ
mod n2
= gmλ
mod n2
⇐ Eq. (1)
= (1 + kn)mλ
mod n2
⇐ Def. of g
= 1 + mλkn mod n2
,
17

where the last equality is derived from (1+an)k ≡ 1+ank (mod n2), which can be
proven by using the binomial theorem. Note that the above operation eliminates
the random number r used in the encryption. Similarly, gλ can be calculated as
gλ
= (1 + kn)λ
mod n2
= 1 + λkn mod n2
.
Thus, by applying function L to them, we have
L(cλ mod n2)
L(gλ mod n2)
mod n =
mλk
λk
mod n = m mod n,
which certainly realizes the decryption.
A.3 Additive Homomorphism
Pailier cryptosystem satisﬁes the additive homomorphism in the sense that the
decryption of the product of two ciphertexts equals to the sum of plaintexts. Let
c1 and c2 be ciphertexts of messages m1 and m2, respectively, and let c := c1 ×
c2 mod n2. Then, since
cλ
= (c1 · c2)λ
mod n2
= (gm1
· rn
1 · gm2
· rn
2 )λ
mod n2
= g(m1+m2)λ
mod n2
= (1 + kn)(m1+m2)λ
mod n2
= 1 + (m1 + m2)λkn mod n2
,
the additive homomorphism certainly follows. It is also known that Pailier cryp-
tosystem satisﬁes a property such that decrypting the ciphertext of m1 to the power
of m2 yields m1 × m2. Let c1 be the ciphertext of m1 snd c := cm2
1 mod n2. Then,
since
cλ
= (gm1
· rn
1 )λm2
mod n2
= g(m1·m2)λ
mod n2
= (1 + kn)(m1·m2)λ
mod n2
= 1 + (m1 · m2)λkn mod n2
,
this property certainly follows.
18

Privacy Preserving Reputation Calculation in P2P Systems with Homomorphic Encryption

More Related Content

What's hot (14)

Similar to Privacy Preserving Reputation Calculation in P2P Systems with Homomorphic Encryption (20)

More from IJCNCJournal (20)

Recently uploaded (20)

Privacy Preserving Reputation Calculation in P2P Systems with Homomorphic Encryption