Enabling Fine-grained Multi-keyword Search Supporting Classified Sub-dictionaries over Encrypted Cloud Data

1545-5971 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TDSC.2015.2406704, IEEE Transactions on Dependable and Secure Computing
1
Enabling Fine-grained Multi-keyword Search
Supporting Classified Sub-dictionaries over
Encrypted Cloud Data
Hongwei Li, Member, IEEE, Yi Yang, Student Member, IEEE, Tom H. Luan, Member, IEEE,
Xiaohui Liang, Student Member, IEEE, Liang Zhou, Member, IEEE, and Xuemin
(Sherman) Shen, Fellow, IEEE
Abstract—Using cloud computing, individuals can store their data on remote servers and allow data access to public users through the
cloud servers. As the outsourced data are likely to contain sensitive privacy information, they are typically encrypted before uploaded to
the cloud. This, however, significantly limits the usability of outsourced data due to the difficulty of searching over the encrypted data. In
this paper, we address this issue by developing the fine-grained multi-keyword search schemes over encrypted cloud data. Our original
contributions are three-fold. First, we introduce the relevance scores and preference factors upon keywords which enable the precise
keyword search and personalized user experience. Second, we develop a practical and very efficient multi-keyword search scheme.
The proposed scheme can support complicated logic search the mixed “AND”, “OR” and “NO” operations of keywords. Third, we further
employ the classified sub-dictionaries technique to achieve better efficiency on index building, trapdoor generating and query. Lastly,
we analyze the security of the proposed schemes in terms of confidentiality of documents, privacy protection of index and trapdoor,
and unlinkability of trapdoor. Through extensive experiments using the real-world dataset, we validate the performance of the proposed
schemes. Both the security analysis and experimental results demonstrate that the proposed schemes can achieve the same security
level comparing to the existing ones and better performance in terms of functionality, query complexity and efficiency.
Index Terms—Searchable encryption, Multi-keyword, Fine-grained, Cloud computing.
!
1 INTRODUCTION
THE cloud computing treats computing as a utility and
leases out the computing and storage capacities to the
public individuals [1], [2], [3]. In such a framework, the
individual can remotely store her data on the cloud server,
namely data outsourcing, and then make the cloud data open
for public access through the cloud server. This represents a
more scalable, low-cost and stable way for public data access
because of the scalability and high efficiency of cloud servers,
and therefore is favorable to small enterprises.
• H. Li and Y. Yang are with the School of Computer Science and
Engineering, University of Electronic Science and Technology of Chi-
na, Chengdu, Sichuan, China (e-mail: hongweili@uestc.edu.cn; yangy-
i.buku@gmail.com).
• H. Li is with State Key Laboratory of Information Security (Institute of
Information Engineering, Chinese Academy of Sciences, Beijing 100093)
(e-mail: hongweili@uestc.edu.cn).
• T. Luan is with the School of Information Technology, Deakin University,
Melbourne, Australia(e-mail: tom.luan@deakin.edu.au).
• X. Liang is with the Department of Computer Science, Dartmouth College,
Hanover, USA(e-mail: Xiaohui.Liang@dartmouth.edu).
• L. Zhou is with the National Key Laboratory of Science and Technology
on Communication, University of Electronic Science and Technology of
China, China(e-mail: lzhou@uestc.edu.cn).
• X. Shen is with the Department of Electrical and Computer En-
gineering, University of Waterloo,Waterloo, Ontario, Canada(e-mail:
sshen@uwaterloo.ca).
Note that the outsourced data may contain sensitive privacy
information. It is often necessary to encrypt the private data
before transmitting the data to the cloud servers [4], [5].
The data encryption, however, would significantly lower the
usability of data due to the difficulty of searching over the
encrypted data [6]. Simply encrypting the data may still
cause other security concerns. For instance, Google Search
uses SSL (Secure Sockets Layer) to encrypt the connection
between search user and Google server when private data,
such as documents and emails, appear in the search results [7].
However, if the search user clicks into another website from
the search results page, that website may be able to identify
the search terms that the user has used.
On addressing above issues, the searchable encryption (e.g.,
[8], [9], [10]) has been recently developed as a fundamental
approach to enable searching over encrypted cloud data,
which proceeds the following operations. Firstly, the data
owner needs to generate several keywords according to the
outsourced data. These keywords are then encrypted and stored
at the cloud server. When a search user needs to access the
outsourced data, it can select some relevant keywords and
send the ciphertext of the selected keywords to the cloud
server. The cloud server then uses the ciphertext to match
the outsourced encrypted keywords, and lastly returns the
matching results to the search user. To achieve the similar
search efficiency and precision over encrypted data as that of
plaintext keyword search, an extensive body of research has
been developed in literature. Wang et al. [11] propose a ranked
keyword search scheme which considers the relevance scores

2
of keywords. Unfortunately, due to using order-preserving
encryption (OPE) [12] to achieve the ranking property, the
proposed scheme cannot achieve unlinkability of trapdoor.
Later, Sun et al. [13] propose a multi-keyword text search
scheme which considers the relevance scores of keywords and
utilizes a multidimensional tree technique to achieve efficient
search query. Yu et al. [14] propose a multi-keyword top-k
retrieval scheme which uses fully homomorphic encryption to
encrypt the index/trapdoor and guarantees high security. Cao
et al. [6] propose a multi-keyword ranked search (MRSE),
which applies coordinate machine as the keyword matching
rule, i.e., return data with the most matching keywords.
Although many search functionalities have been developed
in previous literature towards precise and efficient searchable
encryption, it is still difficult for searchable encryption to
achieve the same user experience as that of the plaintext
search, like Google search. This mainly attributes to following
two issues. Firstly, query with user preferences is very popular
in the plaintext search [15], [16]. It enables personalized search
and can more accurately represent user’s requirements, but has
not been thoroughly studied and supported in the encrypted
data domain. Secondly, to further improve the user’s experi-
ence on searching, an important and fundamental function is
to enable the multi-keyword search with the comprehensive
logic operations, i.e., the “AND”, “OR” and “NO” operations
of keywords. This is fundamental for search users to prune
the searching space and quickly identify the desired data.
Cao et al. [6] propose the coordinate matching search scheme
(MRSE) which can be regarded as a searchable encryption
scheme with “OR” operation. Zhang et al. [17] propose a
conjunctive keyword search scheme which can be regarded as
a searchable encryption scheme with “AND” operation with
the returned documents matching all keywords. However, most
existing proposals can only enable search with single logic
operation, rather than the mixture of multiple logic operations
on keywords, which motivates our work.
In this work, we address above two issues by developing
two Fine-grained Multi-keyword Search (FMS) schemes over
encrypted cloud data. Our original contributions can be sum-
marized in three aspects as follows:
• We introduce the relevance scores and the preference fac-
tors of keywords for searchable encryption. The relevance
scores of keywords can enable more precise returned
results, and the preference factors of keywords represent
the importance of keywords in the search keyword set
specified by search users and correspondingly enables
personalized search to cater to specific user preferences. It
thus further improves the search functionalities and user
experience.
• We realize the “AND”, “OR” and “NO” operations in the
multi-keyword search for searchable encryption. Com-
pared with schemes in [6], [13] and [14], the proposed
scheme can achieve more comprehensive functionality
and lower query complexity.
• We employ the classified sub-dictionaries technique to
enhance the efficiency of the above two schemes. Exten-
sive experiments demonstrate that the enhanced schemes
can achieve better efficiency in terms of index building,
trapdoor generating and query in the comparison with
schemes in [6], [13] and [14].
The remainder of this paper is organized as follows. In
Section 2, we outline the system model, threat model, security
requirements and design goals. In Section 3, we describe
the preliminaries of the proposed schemes. We present the
developed schemes and enhanced schemes in details in Section
4 and Section 5, respectively. Then we carry out the security
analysis and performance evaluation in Section 6 and Section
7, respectively. Section 8 provides a review of the related
works and Section 9 concludes the paper.
2 SYSTEM MODEL, THREAT MODEL
AND SECURITY REQUIREMENTS
2.1 System Model
As shown in Fig. 1, we consider a system consists of three
entities.
• Data owner: The data owner outsources her data to
the cloud for convenient and reliable data access to the
corresponding search users. To protect the data priva-
cy, the data owner encrypts the original data through
symmetric encryption. To improve the search efficien-
cy, the data owner generates some keywords for each
outsourced document. The corresponding index is then
created according to the keywords and a secret key. After
that, the data owner sends the encrypted documents and
the corresponding indexes to the cloud, and sends the
symmetric key and secret key to search users.
• Cloud server: The cloud server is an intermediate entity
which stores the encrypted documents and correspond-
ing indexes that are received from the data owner, and
provides data access and search services to search users.
When a search user sends a keyword trapdoor to the cloud
server, it would return a collection of matching documents
based on certain operations.
• Search user: A search user queries the outsourced docu-
ments from the cloud server with following three steps.
First, the search user receives both the secret key and
symmetric key from the data owner. Second, according
to the search keywords, the search user uses the secret
key to generate trapdoor and sends it to the cloud server.
Last, she receives the matching document collection from
the cloud server and decrypts them with the symmetric
key.
2.2 Threat Model and Security Requirements
In our threat model, the cloud server is assumed to be “honest-
but-curious”, which is the same as most related works on
secure cloud data search [13], [14], [6]. Specifically, the cloud
server honestly follows the designated protocol specification.
However, the cloud server could be “curious” to infer and
analyze data (including index) in its storage and message
flows received during the protocol so as to learn additional
information. we consider two threat models depending on the
information available to the cloud server, which are also used
in [13], [6].

3
Fig. 1. System model
• Known Ciphertext Model: The cloud server can only
know encrypted document collection C and index collec-
tion I, which are both outsourced from the data owner.
• Known Background Model: The cloud server can pos-
sess more knowledge than what can be accessed in
the known ciphertext model, such as the correlation
relationship of trapdoors and the related statistical of
other information, i.e., the cloud server can possess the
statistical information from a known comparable dataset
which bears the similar nature to the targeting dataset.
Similar to [13], [6], we assume search users are trusted
entities, and they share the same symmetric key and secret
key. Search users have pre-existing mutual trust with the
data owner. For ease of illustration, we do not consider
the secure distribution of the symmetric key and the secret
key between the data owner and search users; it can be
achieved through regular authentication and secure channel
establishment protocols based on the prior security context
shared between search users and the data owner [18]. In
addition, to make our presentations more focused, we do
not consider following issues, including the access control
problem on managing decryption capabilities given to users
and the data collection’s updating problem on inserting new
documents, updating existing documents, and deleting existing
documents, are separated issues. The interested readers on
above issues may refer to [6], [5], [10], [19].
Based on the above threat model, we define the security
requirements as follows:
• Confidentiality of documents: The outsourced documents
provided by the data owner are stored in the cloud server.
If they match the search keywords, they are sent to the
search user. Due to the privacy of documents, they should
not be identifiable except by the data owner and the
authorized search users.
• Privacy protection of index and trapdoor: As discussed in
Section 2.1, the index and the trapdoor are created based
on the documents’ keywords and the search keywords,
respectively. If the cloud server identifies the content of
index or trapdoor, and further deduces any association
between keywords and encrypted documents, it may learn
the major subject of a document, even the content of a
short document [20]. Therefore, the content of index and
trapdoor cannot be identified by the cloud server.
• Unlinkability of trapdoor: The documents stored in the
cloud server may be searched many times. The cloud
server should not be able to learn any keyword informa-
tion according to the trapdoors, e.g., to determine two
trapdoors which are originated from the same keywords.
Otherwise, the cloud server can deduce relationship of
trapdoors, and threaten to the privacy of keywords. Hence
the trapdoor generation function should be randomized,
rather than deterministic. Even in case that two search
keyword sets are the same, the trapdoors should be
different.
3 PRELIMINARIES
In this section, we define the notation and review the secure
kNN computation and relevance score, which will serve as the
basis of the proposed schemes.
3.1 Notation
• F—the document collection to be outsourced, denoted as
a set of N documents F = (F1, F2, · · · , FN ).
• C—the encrypted document collection according to F,
denoted as a set of N documents C = (C1, C2, · · · , CN ).
• FID—the identity collection of encrypted documents C,
denoted as FID = (FID1, FID2, · · · , FIDN ).
• W—the keyword dictionary, including m keywords, de-
noted as W = (w1, w2, · · · , wm).
• I—the index stored in the cloud server, which is built
from the keywords of each document, denoted as I =
(I1, I2, · · · , IN ).
• W—the query keyword set generated by a search user,
which is a subset of W.
• TW
—the trapdoor for keyword set W.
• FID—the identity collection of documents returned to
the search user.
• FMS(CS)—the abbreviation of FMS and FMSCS.
3.2 Secure kNN Computation
We adopt the work of Wong et al. [21] as our foundation.
Wong et al. propose a secure k-nearest neighbor (kNN) scheme
which can confidentially encrypt two vectors and compute
Euclidean distance of them. Firstly, the secret key (S, M1, M2)
should be generated. The binary vector S is a splitting indi-
cator to split plaintext vector into two random vectors, which
can confuse the value of plaintext vector. And M1 and M2 are
used to encrypt the split vectors. The correctness and security
of secure kNN computation scheme can be referred to [21].
3.3 Relevance Score
The relevance score between a keyword and a document
represents the frequency that the keyword appears in the
document. It can be used in searchable encryption for returning
ranked results. A prevalent metric for evaluating the relevance
score is TF × IDF, where TF (term frequency) represents
the frequency of a given keyword in a document and IDF

4
(inverse document frequency) represents the importance of
keyword within the whole document collection. Without loss
of generality, we select a widely used expression in [22] to
evaluate the relevance score as
Score(W, Fj) =
∑
w∈W
1
|Fj|
· (1 + lnfj,w) · ln(1 +
N
fw
) (1)
where fj,w denotes the TF of keyword w in document Fj;
fw denotes the number of documents contain keyword w; N
denotes the number of documents in the collection; and |Fj|
denotes the length of Fj, obtained by counting the number of
indexed keywords.
4 PROPOSED SCHEMES
In this section, we firstly propose a variant of the secure kNN
computation scheme, which serves as the basic framework of
our schemes. Furthermore, we describe two variants of our
basic framework and the corresponding functionalities of them
in detail.
4.1 Basic Framework
The secure kNN computation scheme uses Euclidean distance
to select k nearest database records. In this section, we present
a variant of the secure kNN computation scheme to achieve
the searchable encryption property.
4.1.1 Initialization
The data owner randomly generates the secret key K =
(S, M1, M2), where S is a (m+1)-dimensional binary vector,
M1 and M2 are two (m + 1) × (m + 1) invertible matrices,
respectively, and m is the number of keywords in W. Then
the data owner sends (K, sk) to search users through a secure
channel, where sk is the symmetric key used to encrypt
documents outsourced to the cloud server.
4.1.2 Index building
The data owner firstly utilizes symmetric encryption al-
gorithm (e.g., AES) to encrypt the document collection
(F1, F2, · · · , FN ) with the symmetric key sk [23], the encrypt-
ed document collection are denoted as Cj(j = 1, 2, · · · , N).
Then the data owner generates an m-dimensional binary
vector P according to Cj(j = 1, 2, · · · , N), where each
bit P[i] indicates whether the encrypted document contains
the keyword wi, i.e., P[i] = 1 indicates yes and P[i] = 0
indicates no. Then she extends P to a (m + 1)-dimensional
vector P′
, where P′
[m + 1] = 1. The data owner uses
vector S to split P′
into two (m + 1)-dimensional vectors
(pa, pb), where the vector S functions as a splitting indicator.
Namely, if S[i] = 0(i = 1, 2, · · · , m + 1), pa[i] and pb[i]
are both set as P′
[i]; if S[i] = 1(i = 1, 2, · · · , m + 1),
the value of P′
[i] will be randomly split into pa[i] and pb[i]
(P′
[i] = pa[i]+pb[i]). Then, the index of encrypted document
Cj can be calculated as Ij = (paM1, pbM2). Finally, the data
owner sends Cj||FIDj||Ij (j = 1, 2, · · · , N) to the cloud
server.
4.1.3 Trapdoor generating
The search user firstly generates the keyword set W for
searching. Then, she creates a m-dimensional binary vector Q
according to W, where Q[i] indicates whether the i-th keyword
of dictionary wi is in W, i.e., Q[i] = 1 indicates yes and
Q[i] = 0 indicates no. Further, the search user extends Q to
a (m + 1)-dimensional vector Q′
, where Q′
[m + 1] = −s
(the value of −s will be defined in the following schemes
in detail). Next, the search user chooses a random number
r > 0 to generate Q′′
= r · Q′
. Then she splits Q′′
into two
(m + 1) vectors (qa, qb): if S[i] = 0(i = 1, 2, · · · , m + 1),
the value of Q′′
[i] will be randomly split into qa[i] and qb[i];
if S[i] = 1(i = 1, 2, · · · , m + 1), qa[i] and qb[i] are both
set as Q′′
[i]. Thus, the search trapdoor TW
can be generated
as (M−1
1 qa, M−1
2 qb). Then the search user sends TW
to the
cloud server.
4.1.4 Query
With the index Ij(j = 1, 2, · · · , N) and trapdoor TW
, the
cloud server calculates the query result as
Rj = Ij · TW
= (paM1, pbM2) · (M−1
1 qa, M−1
2 qb)
= pa · qa + pb · qb = P′
· Q′′
= rP′
· Q′
= r · (P · Q − s)
(2)
If Rj > 0, the corresponding document identity FIDj will
be returned.
Discussions: The Basic Framework has defined the funda-
mental system structure of the developed schemes. Based on
the secure kNN computation scheme [21], the complementary
random parameter r further enhances the security. Different
values for parameter s and vectors P and Q can lead to new
variants of the Basic Framework. This will be elaborated in
the follows.
4.2 FMS I
In the Basic Framework, P is a m-dimensional binary vector,
and each bit P[i] indicates whether the encrypted document
contains the keyword wi. In the FMS I, the data owner
first calculates the relevance score between the keyword wi
and document Fj. The relevance score can be calculated as
follows:
Score(wi, Fj) =
1
|Fj|
· (1 + lnfj,wi ) · ln(1 +
N
fwi
) (3)
where fj,wi denotes the TF of keyword wi in document Fj;
fwi denotes the number of documents contain keyword wi; N
denotes the number of documents in the collection; and |Fj|
denotes the length of Fj, obtained by counting the number of
indexed keywords.
Then the data owner replaces the value of P[i] with the
corresponding relevance score. On the other hand, we also
consider the preference factors of keywords. The preference
factors of keywords indicate the importance of keywords in
the search keyword set personalized defined by the search
user. For a search user, he may pay more attention to the
preference factors of keywords defined by himself than the
relevance scores of the keywords. Thus, our goal is that

5
if a document has a keyword with larger preference factor
than other documents, it should have a higher priority in
the returned FID; and for two documents, if their largest
preference factor keywords are the same, the document with
higher relevance score of the keyword is the better matching
result.
As shown in Fig. 2, we replace the values of P[i] and
Q[i] by the relevance score and the preference factor of a
keyword, respectively (thus P and Q are no longer binary).
The search user can dynamically adjust the preference factors
to achieve a more ﬂexible search. For convenience, the score
is rounded up, i.e., Score(wi, Fj) = ⌈10 ∗ Score(wi, Fj)⌉,
and we assume the relevance score is not more than D,
i.e., Score(wi, Fj) < D. For the search keyword set W =
(wn1 , wn2 , · · · , wnl
)(1 ≤ n1 < n2 < · · · < nl ≤ m) which
is ordered by ascending importance, the search user randomly
chooses a super-increasing sequence (d1 > 0, d2, · · · , dl) (i.e.,∑j−1
i=1 di ·D < dj(j = 2, 3, · · · , l)), where di is the preference
factor of keyword wni . Then the search result would be:
Rj = r · (P · Q − s) = r · (
l∑
i=1
Score(wni , Fj) · di − s) (4)
Theorem 1: (Correctness) For the search keyword set W =
(wn1 , wn2 , · · · , wnl
)(1 ≤ n1 < n2 < · · · < nl ≤ m) which
is ordered by ascending preference factors, if F1 contains a
larger preference factor keyword compared with F2, then F1
has higher priority in the returned FID.
Proof: For the search keyword set W =
(wn1 , wn2 , · · · , wnl
), assume the keyword sets F1
and F2 contain in W are denoted as W1 =
(wni , · · · , wnx )(n1 ≤ ni < · · · < nx ≤ nl) and
W2 = (wnj , · · · , wny )(n1 ≤ nj < · · · < ny ≤ nl),
respectively, where W1 and W2 are both ordered by
ascending preference factors, and nx > ny. As stated above,
Score(wnx , Fj) ≥ 1 since the score is rounded up, and
∑j−1
i=1 di · D < dj(j = 2, 3, · · · , l). Therefore, there will be
R2 = r · (
∑
wnj
∈W2
Score(wnj , F2) · dj − s)
< r · (
y∑
j=1
< r · (
y∑
j=1
D · dj − s) < r · (dx − s)
< r · (Score(wnx , F1) · dx − s)
< r · (
∑
wni
∈W1
Score(wni , F1) · di − s)
< R1
(5)
Therefore, F1 has higher priority in the returned FID.
Theorem 2: (Correctness) For the search keyword set W =
(wn1 , wn2 , · · · , wnl
)(1 ≤ n1 < n2 < · · · < nl ≤ m)
which is ordered by ascending preference factors, if the largest
preference factor keyword F1 contains is the same as that
F2 contains, and F1 have the higher relevance score of the
keyword, then F1 have higher priority in the returned FID.
Proof: For the search keyword set W =
(wn1 , wn2 , · · · , wnl
), assume the keyword sets F1 and
F2 contain are denoted as W1 = (wni , · · · , wnx )(n1 ≤
ni < · · · < nx ≤ nl) and W2 = (wnj , · · · , wnx )(n1 ≤
nj < · · · < nx ≤ nl), respectively, where W1 and W2
are both ordered by ascending preference factors and
Score(wnx , F1) − Score(wnx , F2) ≥ 1. Thus, there will be
R1 =r · (
∑
wni
∈W1
Score(wni , F1) · di − s)
≥ r · (Score(wnx , F1) · dx − s)
(7)
R2 =r · (
∑
wnj
∈W2
Score(wnj , F2) · dj − s) (8)
=r · (Score(wnx , F2) · dx
+
∑
wnj
∈W2−wnx
<r · (Score(wnx , F2) · dx +
∑
wnj
∈W2−wnx
D · dj − s)
<r · (Score(wnx , F2) · dx + dx − s)
R1 − R2 > r · ((Score(wnx , F1) − Score(wnx , F2)) · dx − dx)
> r · (dx − dx)
> 0 (9)
Therefore, F1 have higher priority in the returned FID than
F2.
Example. We present a concrete example to help under-
stand Theorem 2. The example also illustrates the work-
ing process of FMS I. Speciﬁcally, we assume that the
search keyword set is W = (wn1 , wn2 , · · · , wn5 ), and the
largest preference factor keyword of sets F1 and F2 is the
same, which is wn4 . In addition, we assume the keyword
sets F1 and F2 are W1 = (wn2 , wn3 , wn4 ) and W2 =
(wn1 , wn3 , wn4 ) respectively. Furthermore, we assume that
the relevance score is not more than D = 5, and spe-
cially, let Score(wn4 , F1) = 4 and Score(wn4 , F2) = 2,
which satisfy Score(wn4 , F1) − Score(wn4 , F2) = 2 ≥
1. we randomly choose a super-increasing sequence di =
{1, 10, 60, 500, 3000}(i = 1, · · · , 5), for arbitrary r > 0, there
will be
R1 =r · (
∑
wni
∈W1
Score(wni , F1) · di − s) (11)
≥r · (Score(wn4 , F1) · d4 − s)
≥r · (4 · 500 − s)
≥r · (2000 − s)

6
Data owner Cloud server Search user
1 2
1 2
1 2
1 2 2
2
keyword dictionary: ( , , )
keywords of document : ( , , )
corresponding ( , ): ( , , )
in :( , , , , , , , , )
( 0 ,0 , , , , , , , )
j
i j
w w
F w w
Score w F S S
w w w w w
P S S S
P P
α β
α β
= ⋅⋅⋅
= ⋅⋅⋅
⋅⋅⋅
⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
= ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
′→ →
W
W
W W
1 2
( , )
( , )
a b
j a b
p p
I p M p M=
×
j
I TW
1 2
2
1 2
1 2
1 2
2
search keyword set: ( , , )
super-increasing sequence: ( , , )
in :( , , , , , , , , )
( 0 , 0 , , , , , ,
l
n n n
l
n n n
w w
w w w
d d d
w w w w w
Q d d d
α β
α β
= ⋅⋅⋅
= ⋅⋅⋅
⋅⋅⋅
⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
= ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
W
W
W W
1 1
1 2
, )
( , )
( , )
a b
a b
Q Q r Q q q
T M q M q− −
⋅⋅⋅
′ ′→ → ⋅ →
=W
1
( )
( ( , ) )
larger , better result.
i
j
l
n j ii
j
R r P Q s
r Score w F d s
R
=
= ⋅ ⋅ −
= ⋅ ⋅ −∑
Fig. 2. Structure of the FMS I
R2 =r · (
∑
wnj
∈W2
Score(wnj , F2) · dj − s) (12)
=r · (Score(wn4 , F2) · d4
+
∑
wnj
∈W2−wn4
<r · (Score(wn4 , F2) · dx +
∑
wnj
∈W2−wn4
D · dj − s)
<r · (Score(wn4 , F2) · d4 + d4 − s)
<r · (2 · 500 + 500 − s)
<r · (1500 − s)
R1 − R2 >r · (2000 − s) − r · (1500 − s) (13)
>r · (2000 − 1500)
>500 · r > 0
4.3 FMS II
In the FMS II, we do not change the vector P in the Basic
Framework, but replace the value of Q[i] by the weight of
search keywords, as shown in Fig. 3. With the weight of
keywords, we can also implement some operations like “OR”,
“AND” and “NO” in the Google Search to the searchable
encryption.
Assume that the keyword sets corresponding to the
“OR”, “AND” and “NO” operations are (w′
1, w′
2, · · · , w′
l1
),
(w′′
1 , w′′
2 , · · · , w′′
l2
) and (w′′′
1 , w′′′
2 , · · · , w′′′
l3
), respectively.
Denote “OR”, “AND” and “NO” operations by ∨, ∧ and
¬, respectively. Thus the matching rule can be represented
as (w′
1 ∨ w′
2 ∨ · · · ∨ w′
l1
) ∧ (w′′
1 ∧ w′′
2 ∧ · · · ∧ w′′
l2
) ∧
(¬w′′′
1 ∧ ¬w′′′
2 ∧ · · · ∧ ¬w′′′
l3
). For “OR” operation,
the search user chooses a super-increasing sequence
(a1 > 0, a2, · · · , al1 )(
∑j−1
k=1 ak < aj(j = 2, · · · , l1)) to
achieve searching with keyword weight. To enable searchable
encryption with “AND” and “NO” operations, the search
user chooses a sequence (b1, b2, · · · , bl2 , c1, c2, · · · , cl3 ),
where
∑l1
k=1 ak < bh(h = 1, 2, · · · , l2) and
∑l1
k=1 ak +
∑l2
h=1 bh < ci(i = 1, 2, · · · , l3).
Assume (w′
1, w′
2, · · · , w′
l1
) are ordered by ascending
importance, then according to the search keyword set
(w′
1, w′
2, · · · , w′
l1
, w′′
1 , w′′
2 , · · · , w′′
l2
, w′′′
1 , w′′′
2 , · · · , w′′′
l3
),
the corresponding values in Q are set as
(a1, a2, · · · , al1 , b1, b2, · · · , bl2 , −c1, −c2, · · · , −cl3 ). Other
values in Q are set as 0. Finally, the search user sets
s =
∑l2
h=1 bh. In the Query phase, For a document Fj, if
the corresponding Rj > 0, we claim that Fj can satisfy the
above matching rule.
Theorem 3: (Correctness) Fj satisfies the above matching
rule with “OR”, “AND” and “NO” if and only if the corre-
sponding Rj > 0.
Proof: Firstly, we proof the completeness. Since the weight
of w′′′
i (i = 1, 2, · · · , l3) in the vector Q is −ci and ci >∑l1
k=1 ak +
∑l2
h=1 bh, if any corresponding value of w′′′
i in P
of Fj is 1, we can infer P ·Q < 0 and Rj = r·(P ·Q−s) < 0.
Therefore, if Rj > 0, any of w′′′
i is not in the keyword set of
Fj, i.e., Fj satisfies the “NO” operation. Moreover, if Rj > 0,
then r · (P · Q − s) = r · (P · Q −
∑l2
h=1 bh) > 0. Since
bh >
∑l1
k=1 ak(h = 1, 2, · · · , l2), all corresponding values of
w′′
h in P have to be 1 and at least one corresponding value of
w′
k(k = 1, 2, · · · , l1) in P should be 1. Thus, Fj satisfies the
“AND” and “OR” operations. Therefore, if Rj > 0, the vector
P satisfies the operations of “OR”, “AND” and “NO”.
Next, we show the soundness. If the vector P satisfies
the operations of “OR”, “AND” and “NO”, i.e., at least one
corresponding value of keyword w′
k in P is 1 (assume this
keyword is w′
γ(1 ≤ γ ≤ l1)), all corresponding values of
keywords w′′
h in P are 1 and no corresponding value of
keyword w′′′
i in P is 1. Therefore, Rj = r · (P · Q − s) ≥
r · (aγ + b1 + b2 + · · · + bl2 − s) = r · aγ > 0.
Example. We present a concrete example to help understand
Theorem 3. The example also illustrates the working process
of FMS II. Specifically, we assume that the keyword sets
corresponding to the “OR”, “AND” and “NO” operations are
(w′
1, w′
2, w′
3), (w′′
1 , w′′
2 , w′′
3 ) and (w′′′
1 , w′′′
2 ), respectively. Thus,
the matching rule can be represented as (w′
1 ∨ w′
2 ∨ w′
3) ∧
(w′′
1 ∧ w′′
2 ∧ w′′
3 ) ∧ (¬w′′′
1 ∧ ¬w′′′
2 ). we assume that the search
weights (a1, a2, a3), (b1, b2, b3) and (c1, c2) for “OR”, “AND”
and “NO” are(1,5,8), (20,24,96) and (-500,-600), respectively.
We firstly prove Rj > 0 when Fj satisfies the matching
rule. Specifically, assume that Fj satisfies the matching rule
w′
2 ∧(w′′
1 ∧w′′
2 ∧w′′
3 )∧(¬w′′′
1 ∧¬w′′′
2 ). Thus the corresponding
values of vector P are (0, 1, 0), (1, 1, 1) and (0, 0), respective-
ly. Thus, the result of s =
∑3
h=1 bh = 20 + 24 + 96 = 140,

7
Data owner Cloud server Search user
×
j
I TW1 2
1 2
1 2 2
1 2
keywords of document : ( , , )
in :( , , , , , , , , )
( 0, 0 , , 1 , , 1 , , 1 , )
( , )
( , )
j
a b
j a b
w w
F w w
w w w w w
P
P P p p
I p M p M
α β
= ⋅⋅⋅
= ⋅⋅⋅
⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
= ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
′→ →
=
W
W
W W
1 1
2 2
1 2
1 2 1 2
1 2 1 2
1 2
operation keyword set keyword weight
OR ( , , , ) ( , , , )
AND ( , , , ) ( , , , )
NO ( , , ,
l l
l l
w w
w w w a a a
w w w b b b
w w
= ⋅⋅⋅
′ ′ ′⋅⋅⋅ ⋅⋅⋅
′′ ′′ ′′⋅⋅⋅ ⋅⋅⋅
′′′ ′′′ ⋅⋅⋅
W
3 31 2
1 2
1 1
1 2
) ( , , , )
in : ( , , , , , , , , )
( 0, 0 , , , , , , , )
( , )
( , )
l l
a b
a b
w c c c
w w w w w
Q a c b
Q Q r Q q q
T M q M q
α β χ
α β χ
− −
′′′ ⋅⋅⋅
′ ′′′ ′′⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅
= ⋅⋅⋅ ⋅⋅⋅ − ⋅⋅⋅ ⋅⋅⋅
′ ′→ → ⋅ →
=W
W
( )
( )
0 satisfy "OR","AND" and "NO"
larger larger weight of "OR"
j
j
j
R r P Q s
r a c b s
R
R
α β χ
= ⋅ ⋅ −
= ⋅ ⋅⋅⋅+ +⋅⋅⋅− +⋅⋅⋅+ +⋅⋅⋅−
> ⇒
⇒
Fig. 3. Structure of the FMS II
for arbitrary r > 0, the result of Rj will be
Rj = r · (P · Q − s)
= r · (a2 + b1 + b2 + b3 − s)
= r · (5 + 20 + 24 + 96 − 140)
= 5r > 0
(14)
From the above example, we can easily see that Rj > 0
when Fj satisfies the matching rule. Next, we show that
Rj < 0 when Fj does not satisfy the matching rule. Especially,
we assume that the ”AND” operation does not satisfy the
matching rule. Here, we set the first keyword does not match
the rule, therefore the search keyword set of ”AND” operations
are (0, 1, 1) instead of (1, 1, 1). Thus, the result of Ri will be
Rj = r · (P · Q − s)
= r · (a2 + b2 + b3 − s)
= r · (5 + 24 + 96 − 140)
= −15r < 0
(15)
Obviously, Rj < 0 when Fj does not satisfy the matching
rule.
5 ENHANCED SCHEME
In practice, apart from some common keywords, other key-
words in dictionary are generally professional terms, and this
part of the dictionary will rapidly increase when the dictionary
becomes larger and more comprehensive. Simultaneously, the
data owner’s index will become longer, although many di-
mensions of keywords will never appear in her documents.
That will cause redundant computation and communication
overhead.
In this section, we further propose a Fine-grained Multi-
keyword Search scheme supporting Classified Sub-dictionaries
(FMSCS), which classifies the total dictionary as a common
sub-dictionary and many professional sub-dictionaries. Our
goal is to significantly reduce the computation and commu-
nication overhead. We have researched in a file set randomly
chosen from the National Science Foundation (NSF) Research
Awards Abstracts 1990-2003 [24]. As shown in Fig. 4, we
classify the total dictionary to many sub-dictionaries such
as common sub-dictionary, computer science sub-dictionary,
mathematics sub-dictionary and physics sub-dictionary, etc.
And the search process will only be some minor changes in
Initialization.
Change of Initialization: Compared with the
Basic Framework, in the enhanced scheme the
data owner should firstly choose corresponding sub-
dictionaries. Then her own dictionary can be combined as
{f1||Subdic1||f2||Subdic2|| · · · }, where Subdici represents
all keywords contained in corresponding sub-dictionary and
fi is filling factor with random length which will be 0 string
in the index, the filling factor is used to confuse length of
the data owner’s own dictionary and relative positions of sub-
dictionaries. Then, the data owner and search user will use this
dictionary to generate the index and trapdoor, respectively.
Note that in an dictionary, two professional sub-dictionaries
can even contain a same keyword, but only the first appeared
keyword will be used to generate index and trapdoor, another
will be set to 0 in the vector. And the secret key K will
be formed as (S, M1, M2, |f1|, DID1 , |f2|, DID2 , · · · ), where
DIDi represents the identity of sub-dictionary and |fi| is the
length of fi. Other than these changes, the remaining phases
(i.e., Index building, Trapdoor generating and Query) are
same as the Basic Framework.
Dictionary Updating: In the searchable encryption
schemes with dictionary, dictionary update is a challenge
problem because it may cause to update massive indexes
outsourced to the cloud server. In general dictionary-based
search schemes, e.g., [13] and [14], the update of dictionary
will lead to re-generation of all indexes. In our FMSCS
schemes, when it needs to change the sub-dictionaries or add
new sub-dictionaries, only the data owners who use the corre-
sponding sub-dictionaries need to update their indexes, most
other data owners do not need to do any update operations.
Such dictionary update operations are particularly lightweight.
In addition, Li et al. [9] utilize the dimension expansion
technique to implement the efficient dictionary expansion.
Such method can also be included into our dictionary updating
process. And our scheme can even be more efficient than [9]
since although [9] does not need to re-generate all indexes,
but the corresponding extended operations on all indexes are
necessary. In comparison, our schemes only need to extend
the indexes of partial data owners.

8
RPPRQ RPSXWHU VFLHQFH 0DWKHPDWLFV 3KVLFV ggg
JOREXODU VRIWZDUH QXPEHU PHFKDQLVP ggg
LQYHVWLJDWLRQ VXSHUFRPSXWLQJ VWDWLVWLFDO VWUXFWXUH ggg
WUHH HOHFWULF FROXPQ DHURVRO ggg
UHODWLRQVKLSV YLUWXDO SUREDELOLW SHUFRODWLRQ ggg
ggg ggg ggg ggg ggg
.HZRUGV
6XEGLFWLRQDULHV
1 2 3 4Dictionary: || common || || computer science || || mathematics ||f f f f
1 2 3 4Dictionary: || physics || || mathematics || || common ||f f f f′ ′ ′ ′
ĂƚĂ ŽǁŶĞƌ ϭ
ĂƚĂ ŽǁŶĞƌ Ϯ
Fig. 4. Classified sub-dictionaries
6 SECURITY ANALYSIS
In this section, we analyze the main security properties of
the proposed schemes. In particular, our analysis focuses
on how the proposed schemes can achieve confidentiality
of documents, privacy protection of index and trapdoor, and
unlinkability of trapdoor. Other security features are not the
focus of our concern.
6.1 Confidentiality of Documents
In our schemes, the outsourced documents are encrypted by
the traditional symmetric encryption algorithm (e.g., AES). In
addition, the secret key sk is generated by the data owner and
sent to the search user through a secure channel. Since the
AES encryption algorithm is secure [23], any entity cannot
recover the encrypted documents without the secret key sk.
Therefore, the confidentiality of encrypted documents can be
achieved.
6.2 Privacy Protection of Index and Trapdoor
As shown in Section 4.1, both the index Ij = (paM1, pbM2)
and the trapdoor TW
= (M−1
1 qa, M−1
2 qb) are ciphertexts
of vectors (P, Q). The secret key is K = (S, M1, M2) in
the FMS or (S, M1, M2, |f1|, DID1 , |f2|, DID2 , · · · ) in the
FMSCS, where S functions as a splitting indicator which
splits P and Q into (pa, pb) and (qa, qb), respectively, two
invertible matrices M1 and M2 are used to encrypt (pa, pb)
and (qa, qb). The security of this encryption algorithm has been
proved in the known ciphertext model [21]. Thus, the content
of index and trapdoor cannot be identified. Therefore, privacy
protection of index and trapdoor can be achieved.
6.3 Unlinkability of Trapdoor
To protect the security of search, the unlinkability of trap-
door should be achieved. Although the cloud server cannot
directly recover the keywords, the linkability of trapdoor may
cause leakage of privacy, e.g., the same keyword set may be
searched many times, if the trapdoor generation function is
deterministic, even though the cloud server cannot decrypt
the trapdoors, it can deduce the relationship of keywords. We
consider whether the trapdoor TW
= (M−1
1 qa, M−1
2 qb) can be
linked to the keywords. We prove our schemes can achieve the
unlinkability of trapdoor in a strong threat model, i.e., known
background model [6].
Known Background Model: In this model, the cloud
server can possess the statistical information from a known
comparable dataset which bears the similar nature to the
targeting dataset.
TABLE 1
Structure of Q′
Q′[1] · · · Q′[m] Q′[m + 1]
FMS(CS) I · · · 0 · · · di · · · 0 · · · dj · · · −s
FMS(CS) II · · · ak · · · bh · · · 0 · · · ci · · · −s
As shown in Table 1, in our FMS(CS) I, the trapdoor
is constituted by two parts. The values of all dimensions
di(i = 1, 2, · · · , l) are the super-increasing sequence randomly
chosen by the search user (assume there are α possible
sequences). And the (m + 1) dimension is −s defined by the
search user, where s is a positive random number. Assume
the size of −s is ηs bits, there are 2ηs
possible values for
−s. Further, to generate Q′′
= r · Q′
, Q′
is multiplied by a
positive random number r, there are 2ηr
possible values for
r (if the search user chooses ηr-bit r). Finally, Q′′
is split
to (qa, qb) according the splitting indicator S. Specifically, if
S[i] = 0(i = 1, 2, · · · , m + 1), the value of Q′′
[i] will be
randomly split into qa[i] and qb[i], assume in S the number
of ‘0’ is µ, and each dimension of qa and qb is ηq bits. Note
that ηs, ηr, µ and ηq are independent of each other. Then
in our FMS(CS) I, we can compute the probability that two
trapdoors are the same as follows:
P1 =
1
α · 2ηs · 2ηr · (2ηq )µ
=
1
α · 2ηs+ηr+µηq
(16)
Therefore, the larger α, ηs, ηr, µ and ηq can achieve the
stronger security, e.g., if we choose 1024-bit r, then the

9
probability P1 1/21024
. As a result, the probability that
two trapdoors are the same is negligible.
And in the FMS(CS) II, because −s = −
∑l2
h=1 bh,
its value depends on the weight sequence
(a1, a2, · · · , al1 , b1, b2, · · · , bl2 , c1, c2, · · · , cl3 ). Assume
the number of different sequences is denoted as β, then we
can compute:
P2 =
1
β · 2ηr · (2ηq )µ
=
1
β · 2ηr+µηq
(17)
Similarly, in the FMS(CS) II and the FMS(CS) III, the proba-
bility that two trapdoors are the same is negligible. Therefore,
in our schemes, the unlinkability of trapdoor can be achieved.
In summary, we present the comparison results of security
level in Table 2, where I and II represent FMS(CS) I and
FMS(CS) II, respectively. It can be seen that all schemes can
achieve confidentiality of documents and privacy protection
of index and trapdoor, but the OPE schemes [11], [25] cannot
achieve the unlinkability of trapdoor very well because of the
similarity relevance mentioned in [14].
TABLE 2
Comparison of Security Level
[11], [25] [6], [13], [14] I II
Confidentiality
√ √ √ √
Privacy protection
√ √ √ √
Unlinkability
√ √ √
Discussions: In MRSE [6], the values of P ·Q are equal to
the number of matching keywords, which suffers scale analysis
attack when the cloud server is powerful and has knowledge
of some background information. To solve this problem, it
extends the index and inserts a random number εj which
follows a normal distribution and can confuse the values of
P · Q. Thus, enhanced MRSE can resist scale analysis attack.
However, the introduction of εj causes precision decrease of
the returned results. There is a trade-off between precision
and security in MRSE. In comparison, our schemes do not
suffer the scale analysis attack. Because the values of P · Q
in our schemes do not disclose any information due to the
randomly selected sequences mentioned in Section 4.2 and
Section 4.3. Therefore, our proposal can achieve the security
without sacrificing precision.
7 PERFORMANCE EVALUATION
In this section, we evaluate the performance of the proposed
schemes using simulations, and compare the performance with
that of existing proposals in [6], [13], [14]. We apply a real-
world dataset from the National Science Foundation Research
Awards Abstracts 1990-2003 [24], in which we random select
multiple documents and conduct real-world experiments on an
Intel Core i5 3.2 GHz system.
7.1 Functionality
We compare functionalities between [6], [13], [14] and our
schemes in Table 3, where I and II represent FMS(CS) I and
FMS(CS) II, respectively.
MRSE [6] can achieve multi-keyword search and coordinate
matching using secure kNN computation scheme. And [13]
and [14] considers the relevance scores of keywords. Com-
pared with the other schemes, our FMS(CS) I considers both
the relevance scores and the preference factors of keywords.
Note that if the search user sets all relevance scores and
preference factors of keywords as the same, the FMS(CS) I
degrades to MRSE and the coordinate matching can be
achieved. And in the FMS(CS) II, if the search user sets all
preference factors of “OR” operation keywords as the same,
the FMS(CS) II can also achieve the coordinate matching
of “OR” operation keywords. Particularly, the FMS(CS) II
achieves some fine-grained operations of keyword search,
i.e., “AND”, “OR” and “NO” operations in Google Search,
which are definitely practical and significantly enhance the
functionalities of encrypted keyword search.
TABLE 3
Comparison of Functionalities
[6] [13] [14] I II
Multi-keyword search
√ √ √ √ √
Coordinate matching
√ √ √ √ √
Relevance score
√ √ √
Preference factor
√ √
AND OR NO operations
√
7.2 Query Complexity
In the FMS(CS) II, we can implement “OR”, “AND” and
“NO” operations by defining appropriate weights of keywords,
this scheme provides a more fine-grained search than [6],
[13] and [14]. If the keywords to perform “OR”, “AND” and
“NO” operations are (w′
1, w′
2, · · · , w′
l1
), (w′′
1 , w′′
2 , · · · , w′′
l2
)
and (w′′′
1 , w′′′
2 , · · · , w′′′
l3
), respectively. Our FMS(CS) II can
complete the search with only one query, however, in [6],
[13] and [14], they would complete the search through the
following steps:
• For the “OR” operation of l1 keywords, they need only
one query Query(w′
1, w′
2, · · · , w′
l1
) to return a collection
of documents with the most matching keywords (i.e.,
coordinate matching), which can be denoted as X =
Query(w′
1, w′
2, · · · , w′
l1
).
• For the “AND” operation of l2 keywords, [6], [13]
and [14] cannot generate a query for multiple key-
words to achieve the “AND” operation. Therefore, af-
ter costing l2 queries Query(w′′
i )(i = 1, 2, · · · , l2),
they can do the “AND” operation, and the corre-
sponding document set can be denoted as Y =
Query(w′′
1 )
∩
Query(w′′
2 )
∩
· · ·
∩
Query(w′′
l2
).
• For the “NO” operation of l3 keywords, they need l3
queries Query(w′′′
i )(i = 1, 2, · · · , l3), firstly. Then, the
document set of the “NO” operation can be denoted as
Z = Query(w′′′
1 )
∩
Query(w′′′
2 )
∩
· · ·
∩
Query(w′′′
l3
).
• Finally, the document collection achieved “OR”, “AND”
and “NO” operations can be represented as X
∩
Y
∩
Z.
As shown in Fig. 5a, 5b and 5c, to achieve these operations,
the FMS(CS) II can outperform the existing proposals with
less queries generated.

10
2
4
6
8
10
2
4
6
8
10
0
5
10
15
20
25
Number of AND keywordsNumber of NO keywords
Numberofqueries
[6] [13] [14]
FMS(CS)_II
(a)
2
4
6
8
10
2
4
6
8
10
0
5
10
15
20
Number of OR keywordsNumber of NO keywords
Numberofqueries
[6] [13] [14]
FMS(CS)_II
(b)
2
4
6
8
10
2
4
6
8
10
0
5
10
15
20
Number of AND keywordsNumber of OR keywords
Numberofqueries
[6] [13] [14]
FMS(CS)_II
(c)
Fig. 5. Time for building index. (a) Number of queries
for the different number of “AND” and “NO” keywords
with the same number of “OR” keywords, l1 = 5. (b)
Number of queries for the different number of “OR” and
“NO” keywords with the same number of “AND” keywords,
l2 = 5. (c) Number of queries for the different number of
“AND” and “OR” keywords with the same number of “NO”
keywords, l3 = 5.
7.3 Efficiency
7.3.1 Computation overhead
In order to easily demonstrate our scheme computation over-
head, we analysis our scheme from each phase.
Index building. Note that the Index building phase of [6]
is the same as our FMS II scheme, without calculating the
relevance score. And the Index building phase of the FMS I
is the same as [13], containing the relevance score computing.
Compared with the FMS I, the FMS II does not need to calcu-
late the relevance score. And compared with the computation
cost of building index, the cost of calculating the relevance
score is negligible, we do not distinguish them. Moreover,
in our enhanced schemes (FMSCS), we divide the total
dictionary into 1 common sub-dictionary and 20 professional
sub-dictionaries (assume each data owner averagely chooses 1
common sub-dictionary and 3 professional sub-dictionaries to
generate the index). As shown in Fig. 6, we can see the time for
building index is dominated by both the size of dictionary and
the number of documents. And compared with [6], [13], [14]
and our FMS schemes, the FMSCS schemes largely reduce
the computation overhead.
Trapdoor generating. In Trapdoor generating phase, [6]
and [13] firstly creates a vector according to the search
keyword set W, then encrypts the vector by the secure kNN
computation scheme. And [14] also generates a vector and
uses homomorphic encryption to encrypt each dimension. In
comparison, our FMS I and FMS II schemes should firstly
2000 4000 6000 8000 10000
0
500
1000
1500
2000
Size of dictionary
Time(s)
[6] [13] FMS
[14]
FMSCS
(a)
2000 4000 6000 8000 10000
0
100
200
300
400
500
600
Number of documents
Time(s)
[6] [13] FMS
[14]
FMSCS
(b)
Fig. 6. Time for building index. (a) For the different size of
dictionary with the same number of documents, N=6000.
(b) For the different number of documents with the same
size of dictionary, |W| = 4000.
2000 4000 6000 8000 10000
0
200
400
600
800
1000
1200
Size of dictionary
Time(ms)
[6] [13] FMS
[14]
FMSCS
(a)
10 20 30 40 50
0
50
100
150
200
250
300
Number of query keywords
Time(ms)
[6] [13] FMS
[14]
FMSCS
(b)
Fig. 7. Time for generating trapdoor. (a) For the different
size of dictionary with the same number of query keyword-
s, |W|=20. (b) For the different number of query keywords
with the same size of dictionary, |W| = 4000.
generate a super-increasing sequence and a weight sequence,
respectively. But actually, we can pre-select a corresponding
sequence for each scheme, it can also achieve search process
and privacy. Because even if the vectors are the same for
multiple queries, the trapdoors will be not the same due to
the security of kNN computation scheme. Therefore, the com-
putation cost of [6], [13] and all FMS schemes in Trapdoor
generating phase are the same. As shown in Fig. 7, the time
for generating trapdoor is dominated by the size of dictionary,
instead of the number of query keywords. Hence, our FMSCS
schemes are also very efficient in Trapdoor generating phase.
Query. As [6], [13] and the FMS all adopt the secure kNN
computation scheme, the time for query is the same. The
computation overhead in Query phase, as shown in Fig. 8,
is greatly affected by the size of dictionary and the number
of documents, and almost has no relation to the number of
query keywords. Further we can see, our FMSCS schemes
significantly reduce the computation cost in Query phase.
As [14] needs to encrypt each dimension of index/trapdoor
using full homomorphic encryption, its index/trapdoor size is
enormous. Note that, in Trapdoor generating and Query
phases, the computation overheads are not affected by the
number of query keywords. Thus our FMS and FMSCS
schemes are more efficient compared with some multiple-
keyword search schemes [26], [27], as their cost is linear with
the number of query keywords.

11
2000 4000 6000 8000 10000
0
10
20
30
40
50
60
70
Size of dictionary
Time(s)
[6] [13] FMS
[14]
FMSSCS
(a)
2000 4000 6000 8000 10000
0
10
20
30
40
50
Number of documents
Time(s)
[6] [13] FMS
[14]
FMSCS
(b)
10 20 30 40 50
0
5
10
15
20
25
30
35
Number of query keywords
Time(s)
[6] [13] FMS
[14]
FMSCS
(c)
Fig. 8. Time for query. (a) For the different size of
dictionary with the same number of documents and num-
ber of search keywords, N = 6000, |W| = 20. (b) For
the different number of documents with the same size
of dictionary and number of search keywords, |W| =
4000, |W| = 20. (c) For the different number of search
keyword with the same size of dictionary and number of
documents, N = 6000, |W| = 4000.
7.3.2 Storage overhead
As shown in Table 4, we provide a comparison of storage
overhead among several schemes. Specifically, we evaluate the
storage overhead from three parts: the data owner, the search
user and the cloud server.
According to Table 4, in the FMS, the FMSCS as well as
schemes of [6] and [13], the storage overhead of the data
owner are the same. In these schemes, the data owner preserves
her secret key K = (S, M1, M2) and symmetric key sk
locally, where S is an (m+1)-dimensional vector, M1 and M2
are (m+1)×(m+1) invertible matrices. All elements in S, M1
and M2 are the float number. Since the size of a float number
is 4 bytes, the size of K is 4·(m+1)+8·(m + 1)
2
bytes. We
assume that the size of sk is Ssk that is a constant. Thus, the
total size of storage overhead is 4·(m+1)+8·(m + 1)
2
+Ssk
bytes. However, in [14], the storage overhead of data owner
is λ5
/8 bytes, where the λ is the secure parameter. The
storage overhead is 4GB when we choose λ = 128, which is
popular in a full homomorphic encryption scheme. However,
the storage overhead of the FMS and the FMSCS are almost
763MB when we choose m = 10000, which is large enough
for a search scheme. Therefore, the FMS and the FMSCS are
more efficient than scheme in [14] in terms of the storage
overhead of the data owner.
As shown in Table 4, a search user in the FMS, the FMSCS
as well as the schemes of [6] and [13] preserves the secret key
K = (S, M1, M2) and the symmetric key sk locally. There-
fore, the total storage overhead is 4(m+1)+8(m + 1)
2
+Ssk
bytes. However, in [14], the storage overhead is λ5
/8 + λ2
/8
bytes. The storage overhead is 4GB when we choose λ = 128,
which is popular in a full homomorphic encryption scheme.
However, the storage overhead of the FMS and the FMSCS
are almost 763MB when we choose m = 10000, which is
large enough for a search scheme. Therefore, the FMS and
the FMSCS are more efficient than scheme in [14] in terms
of the storage overhead of the search user.
The cloud server preserves the encrypted documents and the
indexes. The size of encrypted documents in all schemes are
the same, i.e., N·Ds. For the indexes, in the FMS and schemes
in [6] and [13], the storage overhead are 8·(m+1)·N bytes.
In the FMSCS, the storage overhead is 8·ε·(m+1)·N bytes,
where 0 ε 1. When m = 1000 and N = 10000 which
are large enough for a search scheme, the storage overhead of
indexes is about 132MB in the FMSCS. And in schemes of [6]
and [13] as well as the FMS, the size of indexes is 760MB with
the same conditions. In scheme in [14], the storage overhead
of indexes is N · Ds + m · N · (λ/8)
5
bytes, it is 4GB when
we choose λ = 128, which is popular in a full homomorphic
encryption scheme. Therefore, the FMS and the FMSCS are
more efficient than scheme in [14] in terms of the storage
overhead of the cloud server.
7.3.3 Communication overhead
As shown in Table 5, we provide a comparison of com-
munication overhead among several schemes. Specifically,
we consider the communication overhead from three parts:
the communication between the data owner and the cloud
server (abbreviated as D-C), the communication between the
search user and the cloud server (abbreviated as C-S) and the
communication between the data owner and the search user
(abbreviated as D-S).
D-C. In the FMS as well as schemes of [6] and [13], the data
owner needs to send information to cloud server in the form
of Cj||FIDj||Ij (j = 1, 2, · · · , N), where the Cj represents
the encrypted documents, FIDj represents the identity of the
document and Ij represents the index. We assume that the
average size of documents is Ds, thus the size of documents
is N · Ds. We assume the encrypted documents identity FID
is a 10-byte string. Thus, the total size of the identity FID
is 10 · N bytes. The index Ij = (paM1, pbM2) contains two
(m+1)-dimensional vectors. Each dimension is a float number
(the size of each float is 4 bytes). Thus, the total size of index is
8·(m+1)·N bytes. Therefore, the total size of communication
overhead is 8·(m+1)·N+10·N+N·Ds bytes. In the FMSCS,
the total size of communication overhead is 8·ε·(m+1)·N +
10·N+N·Ds bytes. If we choose the ε as 0.2, the size of index
is 1.6 · (m + 1) · N bytes, and the total size of communication
of FMSCS is 1.6·(m+1)·N +10·N +Ds ·N bytes. However,
in [14], the communication overhead is N · Ds + m · N · λ5
/8
bytes, where λ is the secure parameter. If we choose λ = 128
which is popular in a full homomorphic encryption scheme
and m = 1000 and N = 10000 which are large enough for
a search scheme, the FMS and the FMSCS are more efficient
than scheme in [14] in terms of the communication overhead
of D-C.

12
TABLE 4
Comparison of Storage Overhead (Bytes). (m represents the size of dictionary; N represents the number of
documents; Ds represents the average size of each encrypted document; λ represents the secure parameter; ε
represents the decrease rate of dictionary by using our classified sub-dictionaries technology; Ssk represents the size
of symmetric key.)
[14] [6], [13] and FMS FMSCS
Data Owner λ5/8 4 · (m + 1) + 8 · (m + 1)2 + Ssk 4 · (m + 1) + 8 · (m + 1)2 + Ssk
Search User λ5/8 + λ2/8 4 · (m + 1) + 8 · (m + 1)2 + Ssk 4 · (m + 1) + 8 · (m + 1)2 + Ssk
Cloud Server N · Ds + m · N · λ5/8 N · Ds + 8 · (m + 1) · N N · Ds + 8 · ε · (m + 1) · N
TABLE 5
Comparison of Communication Overhead (Bytes). (m represents the size of dictionary; N represents the number of
documents; Ds represents the average size of each encrypted document; T represents the number of returned
documents; λ represents the secure parameter; ε represents the decrease rate of dictionary by using our classified
sub-dictionaries technology; Ssk represents the size of symmetric key.)
[14] [6], [13] and FMS FMSCS
D-C N · Ds + m · N · λ5/8 8 · (m + 1) · N + 10 · N + N · Ds 8 · ε · (m + 1) · N + 10 · N + N · Ds
C-S m · λ5/8 + T · Ds 8 · (m + 1) + T · Ds 8 · ε · (m + 1) + T · Ds
D-S λ5/8 + λ2/8 4 · (m + 1) + 8 · (m + 1)2
+ Ssk 4 · (m + 1) + 8 · (m + 1)2
+ Ssk
C-S. The C-S consists of two phases: Query and Results
returning. In the Query phase, a search user in the FMS as well
as the schemes in [6] and [13] sends the trapdoor to the cloud
server in the form of TW
= (M−1
1 qa, M−1
2 qb), which contains
two (m+1)-dimensional vectors. Thus, the communication
overhead is 8·(m+1) bytes. In the FMSCS, the communication
overhead is 8 · ε · (m + 1)(0 ε 1) bytes. In the Results
returning phase, the cloud server sends the corresponding
result to the search user. The communication overhead of C-
S increases along with the number of returned documents
at this point. We assume that the number of the returned
documents is T, thus, the total communication overhead of
cloud server to search user is T · Ds bytes. Therefore, the
total communication overhead of C-S is 8 · m + T · Ds bytes.
In the FMS as well as the schemes in [6] and [13], the total
communication overhead of C-S is 8 · ε · (m + 1) + T · Ds
bytes. In [14], the total communication overhead of C-S is
m·λ5
/8+T ·Ds bytes. If we choose λ = 128 which is popular
in a full homomorphic encryption scheme and m = 1000 and
N = 10000 which are large enough for a search scheme, the
FMS and the FMSCS are more efficient than scheme in [14]
in terms of the communication overhead of C-S.
D-S. From table 5, we can see that the communication
overhead of the FMS, the FMSCS as well as schemes in
[6] and [13] are the same. In the Initialization phase, the
data owner sends the secret key K = (S, M1, M2) and
symmetric key sk to the search user, where S is an (m + 1)-
dimensional vector, M1 and M2 are (m + 1) × (m + 1)
invertible matrices. Thus, the size of the secret key K is
4 · (m + 1) + 8 · (m + 1)
2
bytes. Therefore, the total size
of communication overhead is 4·(m+1)+8·(m + 1)
2
+Ssk
bytes, where the Ssk represents the size of symmetric key.
However, the communication overhead of scheme in [14] is
λ5
/8+λ2
/8 bytes. The communication overhead is 4GB when
we choose λ = 128, which is popular in a full homomorphic
encryption scheme. However, the communication overhead of
the FMS and the FMSCS are almost 763MB when we choose
m = 10000, which is large enough for a search scheme.
Therefore, the FMS and the FMSCS are more efficient than
scheme in [14] in terms of the communication overhead of
D-S.
8 RELATED WORK
There are mainly two types of searchable encryption in litera-
ture, Searchable Public-key Encryption (SPE) and Searchable
Symmetric Encryption (SSE).
8.1 SPE
SPE is first proposed by Boneh et al. [28], which supports
single keyword search on encrypted data but the computation
overhead is heavy. In the framework of SPE, Boneh et al. [27]
propose conjunctive, subset, and range queries on encrypted
data. Hwang et al. [29] propose a conjunctive keyword scheme
which supports multi-keyword search. Zhang et al. [17] pro-
pose an efficient public key encryption with conjunctive-
subset keywords search. However, these conjunctive keywords
schemes can only return the results which match all the
keywords simultaneously, and cannot rank the returned results.
Qin et al. [30] propose a ranked query scheme which uses
a mask matrix to achieve cost-effectiveness. Yu et al. [14]
propose a multi-keyword top-k retrieval scheme with fully
homomorphic encryption, which can return ranked results and
achieve high security. In general, although SPE allows more
expressive queries than SSE [13], it is less efficient, and
therefore we adopt SPE in the work.
8.2 SSE
The concept of SSE is first developed by Song et al. [8].
Wang et al. [25] develop the ranked keyword search scheme,
which considers the relevance score of a keyword. However,
the above schemes cannot efficiently support multi-keyword
search which is widely used to provide the better experience

13
to the search user. Later, Sun et al. [13] propose a multi-
keyword search scheme which considers the relevance scores
of keywords, and it can achieve efficient query by utilizing
the multidimensional tree technique. A widely adopted multi-
keyword search approach is multi-keyword ranked search
(MRSE) [6]. This approach can return the ranked results of
searching according to the number of matching keywords. Li
et al. [10] utilize the relevance score and k-nearest neigh-
bor techniques to develop an efficient multi-keyword search
scheme that can return the ranked search results based on the
accuracy. Within this framework, they leverage an efficient
index to further improve the search efficiency, and adopt the
blind storage system to conceal access pattern of the search
user. Li et al. [19] also propose an authorized and ranked multi-
keyword search scheme (ARMS) over encrypted cloud data
by leveraging the ciphertext policy attribute-based encryption
(CP-ABE) and SSE techniques. Security analysis demonstrates
that the proposed ARMS scheme can achieve collusion resis-
tance. In this paper, we propose FMS(CS) schemes which not
only support multi-keyword search over encrypted data, but
also achieve the fine-grained keyword search with the function
to investigate the relevance scores and the preference factors of
keywords and, more importantly, the logical rule of keywords.
In addition, with the classified sub-dictionaries, our proposal
is efficient in terms of index building, trapdoor generating and
query.
9 CONCLUSION
In this paper, we have investigated on the fine-grained multi-
keyword search (FMS) issue over encrypted cloud data, and
proposed two FMS schemes. The FMS I includes both the
relevance scores and the preference factors of keywords to
enhance more precise search and better users’ experience,
respectively. The FMS II achieves secure and efficient search
with practical functionality, i.e., “AND”, “OR” and “NO”
operations of keywords. Furthermore, we have proposed the
enhanced schemes supporting classified sub-dictionaries (FM-
SCS) to improve efficiency.
For the future work, we intend to further extend the proposal
to consider the extensibility of the file set and the multi-user
cloud environments. Towards this direction, we have made
some preliminary results on the extensibility [5] and the multi-
user cloud environments [19]. Another interesting topic is to
develop the highly scalable searchable encryption to enable
efficient search on large practical databases.
ACKNOWLEDGMENT
This work is supported by the National Natural Science
Foundation of China under Grants 61472065, 61350110238,
61103207, U1233108, U1333127, and 61272525, the Inter-
national Science and Technology Cooperation and Exchange
Program of Sichuan Province, China under Grant 2014H-
H0029, China Postdoctoral Science Foundation funded project
under Grant 2014M552336, and State Key Laboratory of
Information Security Open Foundation under Grant 2015-MS-
02.
REFERENCES
[1] H. Liang, L. X. Cai, D. Huang, X. Shen, and D. Peng, “An smdp-
based service model for interdomain resource allocation in mobile cloud
networks,” IEEE Transactions on Vehicular Technology, vol. 61, no. 5,
pp. 2222–2232, 2012.
[2] M. M. Mahmoud and X. Shen, “A cloud-based scheme for protecting
source-location privacy against hotspot-locating attack in wireless sensor
networks,” IEEE Transactions on Parallel and Distributed Systems,
vol. 23, no. 10, pp. 1805–1818, 2012.
[3] Q. Shen, X. Liang, X. Shen, X. Lin, and H. Luo, “Exploiting geo-
distributed clouds for e-health monitoring system with minimum service
delay and privacy preservation,” IEEE Journal of Biomedical and Health
Informatics, vol. 18, no. 2, pp. 430–439, 2014.
[4] T. Jung, X. Mao, X. Li, S.-J. Tang, W. Gong, and L. Zhang, “Privacy-
preserving data aggregation without secure channel: multivariate poly-
nomial evaluation,” in Proceedings of INFOCOM. IEEE, 2013, pp.
2634–2642.
[5] Y. Yang, H. Li, W. Liu, H. Yang, and M. Wen, “Secure dynamic
searchable symmetric encryption with constant document update cost,”
in Proceedings of GLOBCOM. IEEE, 2014, to appear.
[6] N. Cao, C. Wang, M. Li, K. Ren, and W. Lou, “Privacy-preserving multi-
keyword ranked search over encrypted cloud data,” IEEE Transactions
on Parallel and Distributed Systems, vol. 25, no. 1, pp. 222–233, 2014.
[7] https://support.google.com/websearch/answer/173733?hl=en.
[8] D. X. Song, D. Wagner, and A. Perrig, “Practical techniques for searches
on encrypted data,” in Proceedings of SP. IEEE, 2000, pp. 44–55.
[9] R. Li, Z. Xu, W. Kang, K. C. Yow, and C.-Z. Xu, “Efficient multi-
keyword ranked query over encrypted data in cloud computing,” Future
Generation Computer Systems, vol. 30, pp. 179–190, 2014.
[10] H. Li, D. Liu, Y. Dai, T. H. Luan, and X. Shen, “Enabling efficient
multi-keyword ranked search over encrypted cloud data through blind
storage,” IEEE Transactions on Emerging Topics in Computing, 2014,
DOI10.1109/TETC.2014.2371239.
[11] C. Wang, N. Cao, J. Li, K. Ren, and W. Lou, “Secure ranked keyword
search over encrypted cloud data,” in Proceedings of ICDCS. IEEE,
2010, pp. 253–262.
[12] A. Boldyreva, N. Chenette, Y. Lee, and A. Oneill, “Order-preserving
symmetric encryption,” in Advances in Cryptology-EUROCRYPT.
Springer, 2009, pp. 224–241.
[13] W. Sun, B. Wang, N. Cao, M. Li, W. Lou, Y. T. Hou, and H. Li,
“Verifiable privacy-preserving multi-keyword text search in the cloud
supporting similarity-based ranking,” IEEE Transactions on Parallel and
Distributed Systems, vol. DOI: 10.1109/TPDS.2013.282, 2013.
[14] J. Yu, P. Lu, Y. Zhu, G. Xue, and M. Li, “Towards secure multi-
keyword top-k retrieval over encrypted cloud data,” IEEE Transactions
on Dependable and Secure Computing, vol. 10, no. 4, pp. 239–250,
2013.
[15] A. Arvanitis and G. Koutrika, “Towards preference-aware relational
databases,” in International Conference on Data Engineering (ICDE).
IEEE, 2012, pp. 426–437.
[16] G. Koutrika, E. Pitoura, and K. Stefanidis, “Preference-based query
personalization,” in Advanced Query Processing. Springer, 2013, pp.
57–81.
[17] B. Zhang and F. Zhang, “An efficient public key encryption with
conjunctive-subset keywords search,” Journal of Network and Computer
Applications, vol. 34, no. 1, pp. 262–267, 2011.
[18] D. Stinson, Cryptography: theory and practice. CRC press, 2006.
[19] H. Li, D. Liu, K. Jia, and X. Lin, “Achieving authorized and ranked
multi-keyword search over encrypted cloud data,” in Proceedings of ICC.
IEEE, 2015, to appear.
[20] S. Zerr, E. Demidova, D. Olmedilla, W. Nejdl, M. Winslett, and
S. Mitra, “Zerber: r-confidential indexing for distributed documents,” in
Proceedings of the 11th international conference on Extending database
technology: Advances in database technology. ACM, 2008, pp. 287–
298.
[21] W. K. Wong, D. W.-l. Cheung, B. Kao, and N. Mamoulis, “Secure
knn computation on encrypted databases,” in Proceedings of SIGMOD
International Conference on Management of data. ACM, 2009, pp.
139–152.
[22] J. Zobel and A. Moffat, “Exploring the similarity space,” in ACM SIGIR
Forum, vol. 32, no. 1. ACM, 1998, pp. 18–34.
[23] N. Ferguson, R. Schroeppel, and D. Whiting, “A simple algebraic
representation of rijndael,” in Selected Areas in Cryptography. Springer,
2001, pp. 103–111.
[24] http://kdd.ics.uci.edu/databases/nsfabs/nsfawards.html.

14
[25] C. Wang, N. Cao, K. Ren, and W. Lou, “Enabling secure and efﬁcient
ranked keyword search over outsourced cloud data,” IEEE Transactions
on Parallel and Distributed Systems, vol. 23, no. 8, pp. 1467–1479,
2012.
[26] P. Golle, J. Staddon, and B. Waters, “Secure conjunctive keyword search
over encrypted data,” in Applied Cryptography and Network Security.
Springer, 2004, pp. 31–45.
[27] D. Boneh and B. Waters, “Conjunctive, subset, and range queries on
encrypted data,” in Theory of cryptography. Springer, 2007, pp. 535–
554.
[28] D. Boneh, G. Di Crescenzo, R. Ostrovsky, and G. Persiano, “Public key
encryption with keyword search,” in Advances in Cryptology–Eurocrypt.
Springer, 2004, pp. 506–522.
[29] Y. Hwang and P. Lee, “Zpublic key encryption with conjunctive keyword
search and its extension to a multi-user system,” in Proceeding of
Pairing. Springer, 2007, pp. 2–22.
[30] Q. Liu, C. C. Tan, J. Wu, and G. Wang, “Efﬁcient information retrieval
for ranked queries in cost-effective cloud environments,” in Proceedings
of INFOCOM. IEEE, 2012, pp. 2581–2585.
Hongwei Li is an Associate Professor with the
School of Computer Science and Engineering,
University of Electronic Science and Technology
of China, China. He received the PhD degree
in computer software and theory from University
of Electronic Science and Technology of Chi-
na, China in 2008. He has worked as a Post-
Doctoral Fellow in Department of Electrical and
Computer Engineering at University of Waterloo
for one year until Oct. 2012. His research inter-
ests include network security, applied cryptog-
raphy, and trusted computing. Dr. Li serves as the Associate Editor of
Peer-to-Peer Networking and Applications, the Guest Editor for Peer
to-Peer Networking and Applications Special Issue on Security and
Privacy of P2P Networks in Emerging Smart City. He also serves on
the technical program committees for many international conferences
such as IEEE INFOCOM, IEEE ICC, IEEE GLOBECOM, IEEE WCNC,
IEEE SmartGridComm, BODYNETS and IEEE DASC. He is a member
of IEEE, a member of China Computer Federation and a member of
China Association for Cryptologic Research.
Yi Yang received his B.S. degree in Network
Engineering from Tianjin University of Science
and Technology (TUST) in 2012. Currently, he is
a master student at the School of Computer Sci-
ence and Engineering, University of Electronic
Science and Technology of China (UESTC), Chi-
na. He serves as the reviewer of Peer-to-Peer
Networking and Application, IEEE INFOCOM,
IEEE ICC, IEEE GLOBECOM, IEEE ICCC, etc.
His research interests include cryptography, and
the secure smart grid.
Tom H. Luan received the B.Sc. degree from
Xian Jiaotong University, China, in 2004, the
M.Phil. degree from Hong Kong University of
Science and Technology, Hong Kong, China, in
2007, and Ph.D. degrees from the University
of Waterloo, Canada, in 2012. Since December
2013, he has been the Lecturer in Mobile and
Apps at the School of Information Technology,
Deakin University, Melbourne Burwood, Australi-
a. His research mainly focuses on vehicular
networking, wireless content distribution, peer-
to-peer networking and mobile cloud computing.
Xiaohui Liang received the B.Sc. degree in
Computer Science and Engineering and the
M.Sc. degree in Computer Software and Theory
from Shanghai Jiao Tong University (SJTU), Chi-
na, in 2006 and 2009, respectively. He is current-
ly working toward a Ph.D. degree in the Depart-
ment of Electrical and Computer Engineering,
University of Waterloo, Canada. His research in-
terests include applied cryptography, and secu-
rity and privacy issues for e-healthcare system,
cloud computing, mobile social networks, and
smart grid.
Liang Zhou is a professor with the National
Key Lab of Science and Technology on Com-
munication at University of Electronic Science
and Technology of China, China. His current
research interests include error control coding,
secure communication and cryptography.
Xuemin (Sherman) Shen is a Professor and
University Research Chair, Department of Elec-
trical and Computer Engineering, University of
Waterloo, Canada. He was the Associate Chair
for Graduate Studies from 2004 to 2008. Dr.
Shen’s research focuses on resource manage-
ment in interconnected wireless/wired networks,
wireless network security, vehicular ad hoc
and sensor networks. Dr. Shen served as the
Technical Program Committee Chair for IEEE
VTC’10 Fall and IEEE Globecom’07. He also
serves/served as the Editor-in-Chief for IEEE Network, Peer-to-Peer
Networking and Application, and IET Communications; a Founding Area
Editor for IEEE Transactions on Wireless Communications; an Asso-
ciate Editor for IEEE Transactions on Vehicular Technology, Computer
Networks. Dr. Shen is a registered Professional Engineer of Ontario,
Canada, an IEEE Fellow, an Engineering Institute of Canada Fellow, a
Canadian Academy of Engineering Fellow, and a Distinguished Lecturer
of IEEE Vehicular Technology Society and Communications Society.

Enabling Fine-grained Multi-keyword Search Supporting Classified Sub-dictionaries over Encrypted Cloud Data

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to Enabling Fine-grained Multi-keyword Search Supporting Classified Sub-dictionaries over Encrypted Cloud Data (20)

Recently uploaded (20)

Enabling Fine-grained Multi-keyword Search Supporting Classified Sub-dictionaries over Encrypted Cloud Data