Ieeepro techno solutions ieee java project - vabks verifiable attribute-based keyword search over outsourced encrypted data

VABKS: Verifiable Attribute-based Keyword Search
over Outsourced Encrypted Data
Qingji Zheng† Shouhuai Xu† Giuseppe Ateniese ‡
† University of Texas at San Antonio, USA
‡ Sapienza University of Rome, Italy and Johns Hopkins University, USA
Abstract—It is common nowadays for data owners to outsource
their data to the cloud. Since the cloud cannot be fully trusted,
the outsourced data should be encrypted. This however brings
a range of problems, such as: How should a data owner grant
search capabilities to the data users? How can the authorized data
users search over a data owner’s outsourced encrypted data?
How can the data users be assured that the cloud faithfully
executed the search operations on their behalf? Motivated by
these questions, we propose a novel cryptographic solution, called
verifiable attribute-based keyword search (VABKS). The solution
allows a data user, whose credentials satisfy a data owner’s access
control policy, to (i) search over the data owner’s outsourced
encrypted data, (ii) outsource the tedious search operations
to the cloud, and (iii) verify whether the cloud has faithfully
executed the search operations. We formally define the security
requirements of VABKS and describe a construction that satisfies
them. Performance evaluation shows that the proposed schemes
are practical and deployable.
I. INTRODUCTION
Cloud computing allows data owners to use massive data
storage and vast computation capabilities at a very low price.
Despite the benefits, data outsourcing deprives data owners of
direct control over their outsourced data. To alleviate concerns,
data owners should encrypt their data before outsourcing to the
cloud. However, encryption can hinder some useful functions
such as searching over the outsourced encrypted data while
enforcing an access control policy. Moreover, it is natural to
outsource the search operations to the cloud, while keeping
the outsourced data private. There is a need to allow the
data users to verify whether the cloud faithfully executed the
search operations or not. To the best of our knowledge, existing
solutions cannot achieve these objectives simultaneously.
A. Our Contributions
We propose a novel cryptographic primitive, called verifi-
able attribute-based keyword search (VABKS). This primitive
allows a data owner to control the search, and use of, its
outsourced encrypted data according to an access control
policy, while allowing the legitimate data users to outsource
the (often costly) search operations to the cloud and verify
whether or not the cloud has faithfully executed the search
operations. In other words, a data user with proper credentials
(corresponding to a data owner’s access control policy) can
(i) search over the data owner’s outsourced encrypted data,
(ii) outsource the search operations to the cloud, and (iii)
verify whether or not the cloud has faithfully executed the
search operations. We formally define the security properties
of VABKS and present a scheme that provably satisfies them.
The scheme is constructed in a modular fashion, by using
attribute-based encryption, bloom filter, digital signature, and
a new building-block we call attribute-based keyword search
(ABKS) that may be of independent value. Experimental
evaluation shows that the VABKS solutions are practical.
B. Related Work
To the best of our knowledge, no existing solution is
adequate for what we want to achieve. In what follows we
briefly review the relevant techniques.
Attribute-Based Encryption (ABE). ABE is a popular
method for enforcing access control policies via cryptographic
means. Basically, this technique allows entities with proper
credentials to decrypt a ciphertext that was encrypted accord-
ing to an access control policy [1]. Depending on how the ac-
cess control policy is enforced, there are two variants: KP-ABE
(key-policy ABE) where the decryption key is associated to
the access control policy [2], and CP-ABE (ciphertext-policy
ABE) where the ciphertext is associated to the access control
policy [3]. ABE has been enriched with various features
(e.g., [4]–[7]). In this paper, we use ABE to construct a new
primitive called attribute-based keyword search (ABKS), by
which keywords are encrypted according to an access control
policy and data users with proper cryptographic credentials can
generate tokens that can be used to search over the outsourced
encrypted data. This effectively prevents a data owner from
knowing the keywords a data user is searching for, while
requiring no interactions between the data users and the data
owners/trusted authorities. This is in contrast to [8], where the
data users interact with the data owners/trusted authorities to
obtain search tokens.
Keyword Search over Encrypted Data. This technique
allows a data owner to generate some tokens that can be used
by a data user to search over the data owner’s encrypted data.
Existing solutions for keyword search over encrypted data
can be classified into two categories: searchable encryption
in the symmetric-key setting (e.g., [9]–[18]) and searchable
encryption in the public-key setting (e.g., [8], [19]–[22]). Sev-
eral variants (e.g., [23]–[26]) have been proposed to support
complex search operations. Moreover, searchable encryption in
the multi-users setting has been investigated as well [12], [27],
where the data owner can enforce an access control policy
by distributing some (stateful) secret keys to the authorized

users. However, all these solutions do not solve the problem we
study, because (i) some of these solutions require interactions
between the data users and the data owners (or a trusted
proxy, such as a trapdoor generation entity [8]) to grant
search capabilities, and (ii) all these solutions (except [18])
assume that the server faithfully executed search operations. In
contrast, our solution allows a data user with proper credentials
to issue search tokens by which the cloud can perform keyword
search operations on behalf of the user, without requiring any
interaction with the data owner. Moreover, the data user can
verify whether or not the cloud has faithfully executed the
keyword search operations. This is true even for the powerful
technique called predicate encryption [28], [29], which does
not offer the desired verifiability.
Verifiable Keyword Search. Recently, verifiable keyword
search solutions have been proposed in [30]–[32], where each
keyword is represented as a root of some polynomial. It is
possible to check whether a keyword is present by evaluating
the polynomial on the keyword and verifying whether the
output is zero or not. However, these approaches work only
when keywords are sent in plaintext to the cloud, and are not
suitable for our purpose because the cloud should not learn
anything about the keywords. It is worth mentioning that the
secure verifiable keyword search in the symmetric-key setting
[18] can be insecure in the public-key setting because the
attacker can infer keywords in question via an off-line keyword
guessing attack (in lieu of the off-line dictionary attack against
passwords).
Paper Organization: Section II reviews some cryptographic
preliminaries. Section III defines ABKS and its security
properties, presents KP-ABKS and CP-ABKS schemes and
analyzes their security properties. Section IV defines VABKS
and its security properties, presents the VABKS construction
and analyzes its security. Section V evaluates the performance
of the ABKS and VABKS schemes. Section VI concludes the
paper.
II. PRELIMINARIES
Let a ← S denote selecting an element a from a set S
uniformly at random, || denote the concatenation operation and
string(S) denote the concatenation of elements of S ordered
by their hash values. Let U = {at1, . . . , atn} be a set of
attributes that are used to specify access control policies.
A. Cryptographic Assumption
Let p be an ℓ-bit prime, and G, GT be cyclic groups of
prime order p with generators g, gT , respectively. Let e be
a bilinear map: e : G × G → GT satisfying: (i) ∀a, b ←
Zp, e(ga
, gb
) = e(g, g)ab
, (ii) e(g, g) = 1, and (iii) e can be
computed efficiently.
Decisional Linear Assumption (DL). Given (g, f, h, fr1
,
gr2
, Q) where g, f, h, Q ← G, r1, r2 ← Zp, this assumption
says that any probabilistic polynomial-time algorithm A can
determine Q
?
= hr1+r2
at most with a negligible advantage in
security parameter ℓ, where “advantage” is defined as
|Pr[A(g, f, h, fr1
, gr2
, hr1+r2
) = 1] −
Pr[A(g, f, h, fr1
, gr2
, Q) = 1]|.
Generic Bilinear Group [33]. Let ψ0, ψ1 be two random
encodings of the additive group Z+
p , such that ψ0, ψ1 are
injective maps from Z+
p to {0, 1}m
, where m > 3 log(p). Let
G = {ψ0(x)|x ∈ Zp} and GT = {ψ1(x)|x ∈ Zp}. There is
an oracle to compute e : G × G → GT . G is referred to as a
generic bilinear group. Let g denote ψ0(1), gx
denote ψ0(x),
e(g, g) denote ψ1(1), and e(g, g)y
denote ψ1(y).
Pseudorandom Generator [34]. A pseudorandom genera-
tor H : {0, 1}ℓ
→ {0, 1}m
, ℓ < m, is a deterministic algorithm
that takes as input an ℓ-bit seed and generates a m-bit string
that cannot be distinguished from a m-bit random string by
any polynormial-time algorithm (in ℓ).
B. Bloom Filter for Membership Query
A Bloom filter [35] is a data structure for succinctly
representing a static set, while allowing membership queries.
A m-bit Bloom filter is an array of m bits, which are all
initialized as 0. It uses k independent universal hash functions
H′
1, . . . , H′
k with the same range {0, . . . , m − 1}. For each
element w ∈ S = {w1, . . . , wn}, the bits corresponding
to H′
j(w) are set to 1, where 1 ≤ j ≤ k. To determine
whether w belongs to S or not, once can check whether
all of the bits corresponding to H′
j(w) equal to 1, where
1 ≤ j ≤ k. If not, it is certain that w ∈ S; otherwise,
w ∈ S with a high probability (i.e., there is a non-zero false-
positive rate). Suppose the hash functions are perfectly random
and n elements are hashed into a m-bit Bloom filter, the
false-positive rate is (1 − (1 − 1
m )kn
)k
≈ (1 − e−kn/m
)k
.
Note that k = (ln 2)m/n hash functions lead to the minimal
false-positive rate (0.6185)m/n
. A m-bit Bloom filter has two
associated algorithms:
• BF ← BFGen({H′
1, . . . , H′
k}, {w1, . . . , wn}): This algo-
rithm generates a m-bit Bloom filter by hashing a data
set S = {w1, . . . , wn} with {H′
1, . . . , H′
k}.
• {0, 1} ← BFVerify({H′
1, . . . , H′
k}, BF, w): This algo-
rithm returns 1 if w ∈ S, and 0 otherwise.
C. Access Trees for Representing Access Control Policies
Access trees can represent access control policies [2]. In an
access tree, a leaf is associated with an attribute and an inner
node represents a threshold gate. Let numv be the number of
children of node v, and label the children from the left to the
right as 1, . . . , numv. Let kv, 1 ≤ kv ≤ numv, be the threshold
value associated with node v, where kv = 1 represents the OR
gate and kv = numv represents the AND gate. Let parent(v)
denote the parent of node v, ind(v) denote the label of node
v, att(v) denote the attribute associated to leaf node v, lvs(T)
denote the set of leaves of access tree T, and Tv denote the
subtree of T rooted at node v (e.g., Troot = T).
Let F(Atts, Tv) = 1 indicate that an attribute set Atts
satisfies the access control policy represented by subtree Tv,
where F(Atts, Tv) can be evaluated iteratively as follows:

• In the case v is a leaf: If att(v) ∈ Atts, set F(Atts, Tv) =
1; otherwise, set F(Atts, Tv) = 0.
• In the case v is an inner node with children v1, . . . , vnumv
:
If there exists a subset I ⊆ {1, . . ., numv} such that |I| ≥
kv and ∀ j ∈ I, F(Atts, Tvj ) = 1, set F(Atts, Tv) = 1;
otherwise, set F(Atts, Tv) = 0.
Given an access tree T, we denote the algorithm for
distributing a secret s according to T by:
{qv(0)|v ∈ lvs(T)} ← Share(T, s).
This algorithm generates a polynomial qv of degree kv − 1
for each node v in a top-down fashion (for each leaf node
kv = 1):
• If v is the root of T (i.e., v = root), set qv(0) = s and
randomly pick kv − 1 coefficients for polynomial qv.
• If v is a leaf of T, set qv(0) = qparent(v)(ind(v)).
• If v is an inner node (but not the root), set qv(0) =
qparent(v)(ind(v)) and randomly select kv − 1 coefficients
for polynomial qv.
When the algorithm halts, each leaf v is associated with a
value qv(0), which is the secret share of s at node v.
Given an access tree T and a set of values {Eu1 ,
. . ., Eum }, where u1, . . ., um are the leaves of T,
F({att(u1), . . . , att(um)}, T) = 1, Euj = e(g, h)quj
(0)
for 1 ≤ j ≤ m, g, h ∈ G, e is a bilinear map, and
qu1 (0), . . . , qum (0) are secret shares of s according to T, the
algorithm for reconstructing e(g, h)s
is denoted by
e(g, h)s
← Combine(T, {Eu1 , . . . , Eum }).
This algorithm executes the following steps with respect to
node v in a bottom-top fashion according to T:
• If F({att(u1), . . . , att(um)}, Tv) = 0, then set Ev = ⊥.
• If F({att(u1), . . . , att(um)}, Tv) = 1, then execute the
following:
– If v is a leaf, set Ev = Euj (0) = e(g, h)quj
(0)
where
v = uj for some j.
– If v is an inner node (including the root), for v’s
children nodes {v1, · · · , vnumv
}, there exists a set
of indices S such that |S| = kv, j ∈ S, and
F({att(u1), . . . , att(um)}, Tvj ) = 1. Set
Ev =
j∈S
E
∆vj
vj =
j∈S
(e(g, h)qvj
(0)
)∆vj = e(g, h)qv(0)
,
where ∆vj = l∈S,l=j
−j
l−j .
When the algorithm halts, the root of T is associated with
Eroot = e(g, h)qroot(0)
= e(g, h)s
.
III. ATTRIBUTE-BASED KEYWORD SEARCH (ABKS)
This new primitive allows a data owner to specify a policy
for controlling the keyword search operations over its out-
soured encrypted data. That is, a data user who possesses
attributes that satisfy the data owner’s policy can conduct
keyword search over the oursourced encrypted data. This
primitive naturally has two variants: KP-ABKS (key-policy
ABKS) where the cryptographic credentials are associated to
the access control policy, and CP-ABKS (ciphertext-policy
ABKS) where the ciphertext is associated to the access control
policy. To unify the presentation, let IEnc denote the input
to encryption function Enc and IKeyGen denote the input to
key generation function KeyGen. For CP-ABKS, IEnc and
IKeyGen are respectively the access tree and the attribute set; for
KP-ABKS, IEnc and IKeyGen are respectively the attribute set
and the access tree. Let F(IKeyGen, IEnc) = 1 denote IKeyGen
satisfies IEnc in CP-ABKS and IEnc satisfies IKeyGen in KP-
ABKS.
A. Definition and Security
The model of ABKS is: A data owner outsources its
encrypted keywords to the cloud, a data user generates search
tokens according to some keywords, and the cloud, who
receives search tokens from the user, conducts the search
operations over outsourced encrypted keywords.
Definition 1: ABKS consists of the following algorithms:
• (mk, pm) ← Setup(1ℓ
): This algorithm initializes the
public parameter pm and generates a master key mk.
• sk ← KeyGen(mk, IKeyGen): This algorithm outputs cre-
dential sk for a user according to IKeyGen.
• cph ← Enc(w, IEnc): This algorithm encrypts keyword w
to obtain ciphertext cph.
• tk ← TokenGen(sk, w): This algorithm allows a data user
to generate a search token tk according to its credential
sk and keyword w.
• {0, 1} ← Search(cph, tk): This algorithm returns 1 if (i)
F(IKeyGen, IEnc) = 1 and (ii) ciphertext cph and token tk
correspond to the same keyword , and return 0 otherwise.
An ABKS scheme is correct if the following holds: Given
(mk, pm) ← Setup(1ℓ
), sk ← KeyGen(mk, IKeyGen) and
F(IKeyGen, IEnc) = 1, cph ← Enc(w, IEnc) and tk ←
TokenGen(sk, w), Search(cph, tk) always returns 1.
The adversary model against ABKS is the following: data
owners and authorized data users are trusted, but the cloud is
trusted but curious (i.e., executing the protocol honestly but
attempting to infer private information as well). Intuitively,
security means that the cloud learn nothing beyond the search
results. Specifically, given a probabilistic polynomial-time
adversary A (modeling the cloud), an ABKS scheme is secure
if the following holds
• Selective security against chosen-keyword attack: Without
being given any matching search token, A cannot infer
any information about the plaintext keyword of a keyword
ciphertext in the selective security model, where A must
determine IEnc it intends to attack before the system is
boostrapped [36]. We formalize this security property via
the selective chosen-keyword attack game .
• Keyword secrecy: In the public-key setting, it is impos-
sible to protect the search tokens (aka. predicate privacy
[37]) against the keyword guessing attack. This is because
A can encrypt a keyword of its choice and check whether
the resulting keyword ciphertext and the target token

correspond to the same keyword, which is caused by the
use of “deterministic encryption.” Therefore, we use a
weaker security notion called keyword secrecy, assuring
that the probability A learning the keyword from the
keyword ciphertext and search tokens is negligibly more
than the probability of correct random keyword guess. We
formalize this security property via the keyword secrecy
game.
Selectively Chosen-Keyword Attack (SCKA) Game:
Setup: A selects a non-trivial challenge I∗
Enc (a trivial chal-
lenge I∗
Enc is one that can be satisfied by any data user who
does not have any credential), and gives it to the challenger.
Then the challenger runs Setup(1ℓ
) to generate the public
parameter pm and the master key mk.
Phase 1: A can query the following oracles for polynomially
many times, and the challenger keeps a keyword list Lkw,
which is initially empty.
• OKeyGen(IKeyGen): If F(IKeyGen, I∗
Enc) = 1, then abort;
otherwise, the challenger returns to A credential sk
corresponding to IKeyGen.
• OTokenGen(IKeyGen, w) : The challenger generates creden-
tial sk with IKeyGen, and returns to A a search token tk
by running algorithm TokenGen with inputs sk and w. If
F(IKeyGen, I∗
Enc) = 1, the challenger adds w to Lkw.
Challenge phase: A chooses two keywords w0 and w1, where
w0, w1 /∈ Lkw. The challenger selects λ ← {0, 1}, computes
cph∗
← Enc(wλ, I∗
Enc), and delivers cph∗
to A. Note that the
requirement of w0, w1 /∈ Lkw is to prevent A from trivially
guessing λ with tokens from OTokenGen.
Phase 2: A continues to query the oracles as in Phase 1. The
restriction is that (IKeyGen, w0) and (IKeyGen, w1) cannot be the
input to OTokenGen if F(IKeyGen, I∗
Enc) = 1.
Guess: A outputs a bit λ′
, and wins the game if λ′
= λ.
Let |Pr[λ = λ′
] − 1
2 | be the advantage of A winning the
above SCKA game. Thus, we have
Definition 2: An ABKS scheme is selectively secure against
chosen-keyword attack if the advantage of any A winning the
SCKA game is negligible in security parameter ℓ.
Keyword Secrecy Game:
Setup: The challenger runs Setup(1ℓ
) to generate the public
parameter pm and the master key mk.
many times:
• OKeyGen(IKeyGen): The challenger returns to A credential
sk corresponding to IKeyGen. It adds IKeyGen to the list
LKeyGen, which is initially empty.
• OTokenGen(IKeyGen, w): The challenger generates creden-
by running algorithm TokenGen with input sk and w.
Challenge phase: A chooses a non-trivial I∗
Enc and gives
it to the challenger. The challenger selects w∗
from the
message space uniformly at random and selects I∗
KeyGen such
that F(I∗
KeyGen, I∗
Enc) = 1. The challenger runs cph ←
Enc(w∗
, I∗
Enc) and tk ← TokenGen(sk, w∗
) and delivers
(cph, tk) to A. We require that ∀IKeyGen ∈ LKeyGen,
F(IKeyGen, I∗
Enc) = 0.
Guess: After guessing q distinct keywords, A outputs a
keyword w′
, and wins the game if w′
= w.
Definition 3: An ABKS scheme achieves keyword secrecy
if the probability that A wins the keyword secrecy game is
at most 1
|M|−q + ǫ, where M is the keyword space, q is the
number of distinct keywords that the adversary has attempted,
and ǫ is a negligible in security parameter ℓ.
B. Construction
The basic idea underlying the construction is the following:
each keyword ciphertext and each search token has two parts,
one is associated to the keyword and the other is associated
to the attributes (or access control policy). If the attributes
satisfy the access control policy, one can determine whether
the search token and keyword ciphertext correspond to the
same keyword or not. Consider KP-ABKS as an example. Let
H1 : {0, 1}∗
→ G be a hash function modeled as random
oracle and H2 : {0, 1}∗
→ Zp be an one-way hash function.
A data user’s credentials are generated by letting t ← Zp,
Av = gqv(0)
H1(att(v))t
, Bv = gt
for each leaf v, where g
is a generator of G, qv(0) is the share of secret ac for leaf v
according to access tree T. The keyword ciphertext and search
token are generated as follows:
• Keyword w is encrypted into two parts: one is to “blend”
w with randomness r1, r2 ← Zp by letting W′
=
gcr1
, W = ga(r1+r2)
gbH2(w)r1
and W0 = gr2
where
ga
, gb
, gc
∈ G are public keys, and the other is associated
to attribute set Atts by letting Wj = H1(atj)r2
for each
atj ∈ Atts. The two parts are tied together via r2.
• Given a set of credentials, a search token for keyword
w is generated with two parts: one is associated to w as
tok1 = (ga
gbH2(w)
)s
and tok2 = gcs
for some s ← Zp,
and the other is associated to the credentials by letting
A′
v = As
v, B′
v = Bs
v for each v ∈ lvs(T). The two parts
are tied together via randomness s.
If the attribute set Atts satisfies the access tree T, the cloud
can use A′
v, B′
v and W0, Wj to recover e(g, g)acr2s
, which can
be used to test the keyword equality as elaborated below.
1) KP-ABKS Construction and Security Analysis: Let ℓ be
the primary security parameter. It consists of the following
algorithms.
Setup(1ℓ
): Select a bilinear map e : G×G → GT , where G
and GT are cyclic groups of order p, which is an ℓ-bit prime.
Let H1 : {0, 1}∗
→ G be a hash function modeled as random
oracle and H2 : {0, 1}∗
→ Zp be an one-way hash function,
select a, b, c ← Zp and g ← G, and set
pm = (H1, H2, e, g, p, ga
, gb
, gc
, G, GT ), mk = (a, b, c).
KeyGen(mk, T): Execute Share(T, ac) to obtain secret share
qv(0) of ac for each leave v ∈ lvs(T) on access tree T.
For each leaf v ∈ lvs(T), pick t ← Zp, and compute
Av = gqv(0)
H1(att(v))t
and Bv = gt
. Set
sk = (T, {(Av, Bv)|v ∈ lvs(T)}).

Enc(w, Atts): Select r1, r2 ← Zp, and compute W′
= gcr1
,
W = ga(r1+r2)
gbH2(w)r1
and W0 = gr2
. For each atj ∈ Atts,
compute Wj = H1(atj)r2
. Set
cph = (Atts, W′
, W, W0, {Wj|atj ∈ Atts}).
TokenGen(sk, w): Select s ← Zp, and compute A′
v =
As
v, B′
v = Bs
v for each v ∈ lvs(T). Compute tok1 =
(ga
gbH2(w)
)s
and tok2 = gcs
. Set
tk = (tok1, tok2, T, {(A′
v, B′
v)|v ∈ lvs(T)})
Search(tk, cph): Given attribute set Atts specified in cph,
select an attribute set S satisfying the access tree T spec-
ified in tk. If S does not exist, return 0; otherwise, for
each atj ∈ S, compute Ev = e(A′
v, W0)/e(B′
v, Wj) =
e(g, g)sr2qv(0)
, where att(v) = atj for v ∈ lvs(T). Com-
pute e(g, g)sr2qroot(0)
← Combine(T, {Ev|att(v) ∈ S}) so
that Eroot = e(g, g)acsr2
. Return 1 if e(W′
, tok1)Eroot =
e(W, tok2), and 0 otherwise.
The scheme is correct because
e(W′
, tok1)Eroot = e(gcr1
, (ga
gbH2(w)
)s
)Eroot
= e(g, g)acs(r1+r2)
e(g, g)bcsH2(w)r1
,
e(W, tok2) = e(ga(r1+r2)
gbH2(w)r1
, gcs
)
= e(g, g)acs(r1+r2)
e(g, g)bcsH2(w)r1
The scheme is secure because of the following theorems,
whose proofs are given in Appendix A and Appendix B,
respectively.
Theorem 1: Given the DL assumption and one-way hash
function H2, the KP-ABKS scheme is selectively secure
against chosen-keyword attack in the random oracle model.
Theorem 2: Given the one-way hash function H2, the KP-
ABKS scheme achieves keyword secrecy in the random oracle
model.
2) CP-ABKS Construction and Security Analysis: Let ℓ be
the primary security parameter. It consists of the following
algorithms.
Setup(1ℓ
): Select a bilinear group e : G × G → GT , where
G and GT are cyclic groups of order p, which is an ℓ-bit
prime. Let H1 : {0, 1}∗
→ G be a hash function modeled as
random oracle and H2 : {0, 1}∗
→ Zp be an one-way hash
function, select a, b, c ← Zp and g ← G, and set
pm = (H1, H2, e, g, p, ga
, gb
, gc
, G, GT ), mk = (a, b, c).
KeyGen(mk, Atts): Select r ← Zp, compute A = g(ac−r)/b
.
For each atj ∈ Atts, select rj ← Zp and computes Aj =
gr
H1(atj)rj
and Bj = grj
. Set
sk = (Atts, A, {(Aj, Bj)|atj ∈ Atts}).
Enc(w, T): Select r1, r2 ← Zp, and compute W = gcr1
,
W0 = ga(r1+r2)
gbH2(w)r1
and W′
= gbr2
. Compute secret
shares of r2 for each leave of access tree T as {qv(0)|v ∈
lvs(T)} ← Share(T, r2). For each v ∈ lvs(T), compute Wv =
gqv(0)
and Dv = H1(att(v))qv (0)
. Set
cph = (T, W, W0, W′
, {(Wv, Dv)|v ∈ lvs(T)}).
TokenGen(sk, w): Select s ← Zp, and compute tok1 =
(ga
gbH2(w)
)s
, tok2 = gcs
and tok3 = As
= g(acs−rs)/b
. For
each atj ∈ Atts, compute A′
j = As
j and B′
j = Bs
j . Set
tk = (Atts, tok1, tok2, tok3, {(A′
j, B′
j)|atj ∈ Atts}).
Search(tk, cph): Given attribute set Atts as specified in
tk, select an attribute set S that satisfies the access tree T
specified in cph. If S does not exist, return 0; otherwise,
for each atj ∈ S, compute Ev = e(A′
j, Wv)/e(B′
j, Dv) =
e(g, g)rsqv(0)
, where att(v) = atj for v ∈ lvs(T). Com-
pute e(g, g)rsqroot(0)
← Combine(T, {Ev|att(v) ∈ S})
and Eroot = e(g, g)rsr2
. Return 1 if e(W0, tok2) =
e(W, tok1)Eroote(tok3, W′
), and 0 otherwise.
Correctness of the scheme can be verified similarly to that of
KP-ABKS. Security of the scheme is assured by the following
theorems, The proof of the former one is deferred to Appendix
C, and the proof of the latter one is omitted because it is similar
to that of Theorem 2.
Theorem 3: Given the one-way hash function H2, the CP-
ABKS scheme is selectively secure against chosen-keyword
attack in the generic bilinear group model [33].
Theorem 4: Given the one-way hash function H2, the CP-
ABKS scheme achieves keyword secrecy in the random oracle
model.
IV. VERIFIABLE ATTRIBUTE-BASED KEYWORD SEARCH
In the model of ABKS, the party (e.g., cloud) is assumed to
execute the search operation faithfully (despite that the party
may attempt to infer useful information about the keywords).
VABKS achieves the goal of ABKS despite that the party
executing the search operation may be malicious.
A. Model
We consider the system model illustrated in Figure 1,
which involves four parties: a data owner, who outsources
its encrypted data as well as encrypted keyword-index to
the cloud; a cloud, which provides storage services and can
conduct keyword search operations on behalf of the data users;
a data user, who is to retrieve the data owner’s encrypted data
according to some keyword (i.e., keyword search); a trusted
authority, which issues credentials to the data owners/users.
The credentials are sent over authenticated private channels
(which can be achieved through another layer of mechanisms).
Data owner
Cloud
Search Token for KeywordX
Outsourcing
Datauser (W/ or W/o credentials
satisfying data owner’s access control policies)
Search result & proofTrusted authority
(issuing credentials
for all cloud users)
Y
X
F2
F1
F3
W
V
Keywords Data files
Fig. 1. VABKS system model, where keywords X, Y and V, W may
correspond to different access control policies.

The data owners are naturally trusted. Both authorized and
unauthorized data users are semi-trusted, meaning that they
may try to infer some sensitive information of interest. The
cloud is not trusted as it may manipulate the search operations,
which already implies that the cloud may manipulate the
outsourced encrypted data.
B. Definition
Let FS = {F1, . . . , Fn} be a set of data files. Let KGj,
1 ≤ j ≤ l, be a set of keywords (also called “keyword group”)
that are encrypted with the same access control policy (i.e.,
access tree). Let KG = {KG1, . . . , KGl}. For each keyword w,
let MP(w) be the set of identifiers identifying data files that
contain keyword w. Let MP = {MP(w)|w ∈ ∪l
i=1KGi}. Let
D = (KG, MP, FS) denote keyword-index and the data files.
Definition 4: A VABKS scheme consists of the following
algorithms:
• (mk, pm) ← Init(1ℓ
): This algorithm is run by the trusted
authority to initialize the system.
• sk ← KeyGen(mk, IKeyGen): This algorithm is run by
the trusted authority to issue credentials sk for data
users/owners.
• (Au, Index, Dcph) ← BuildIndex({IEnc}l, {I′
Enc}n, D):
This algorithm is run by a data owner to encrypt D =
(KG, MP, FS) to data ciphertext Dcph, index ciphertext
Index and auxiliary information Au, where {IEnc}l is the
set of access control policies respectively for encrypting
the l keyword groups KG1, . . . , KGl and {I′
Enc}n is the
set of access control policies respectively for encrypting
the n data files FS1, . . . , FSn (It may happen that the
access control policies for keywords and their respective
data files are different).
• tk ← TokenGen(sk, w): This algorithm is run by an
authorized data user to generate a search token tk for
keyword w.
• (proof, rslt) ← SearchIndex(Au, Index, Dcph, tk): This
algorithm is run by the cloud to conduct the search
operations over encrypted index Index on behalf of a data
user. It outputs the search result rslt and a proof proof.
• {0, 1} ← Verify(sk, w, tk, rslt, proof): This algorithm is
run by the data user to verify that (rslt, proof) is valid
with respect to search token tk.
A VABKS scheme is correct if the following holds:
given (mk, pm) ← Init(1ℓ
), sk ← KeyGen(mk, IKeyGen),
(Au, Index, Dcph) ← BuildIndex({IEnc}l, {I′
Enc}n, D), tk ←
TokenGen(sk, w) and (proof, rslt) ← SearchIndex(Au, Index,
Dcph, tk), Verify(sk, w, tk, rslt, proof) always returns 1.
Informally, security of VABKS is defined as the following
four requirements, where the cloud is the adversary A.
• Data secrecy: Given encrypted keywords and search
tokens, A still cannot learn any information (in a compu-
tational sense) about the encrypted data files. This defi-
nition can be formalized by the chosen-plaintext security
game, where two challenges D0 = (KG, MP, FS0), D1 =
(KG, MP, FS1) correspond to the same KG and MP, and
|FS0| = |FS1|.
• Selective security against chosen-keyword attack: Without
seeing corresponding search tokens, A cannot infer any
information about the keyword from the keyword cipher-
text. This property is extended from the selective security
against chosen-keyword attack of ABKS.
• Keyword secrecy: Given encrypted data files, the probabil-
ity that A learn the plaintext keyword from the keyword
ciphertext as well as the search tokens is no more than
that of a random guess. This property is extended from
the keyword secrecy of ABKS.
• Verifiability: If A returns an incorrect search result, it can
be detected by the user with an overwhelming probability.
We formalize this security property via the following
verifiability game.
Verifiability Game:
Setup: The challenger runs (pm, mk) ← Init(1ℓ
). A selects
D = (KG, MP, FS), {IEnc}l and {I′
Enc}n and sends them
to the challenger. The challenger runs (Au, Index, Dcph) ←
BuildIndex({IEnc}l, {I′
Enc}n, D), and gives (Au, Index, Dcph)
to A.
many times.
• OKeyGen(IKeyGen): The challenger returns to A credential
sk corresponding to IKeyGen.
• OTokenGen(IKeyGen, w): The challenger generates creden-
by running algorithm TokenGen with inputs sk and w.
• OVerify(IKeyGen, w, tk, rslt, proof): The challenger gener-
ates credential sk with IKeyGen, returns γ to A by running
γ ← Verify(sk, w, tk, rslt, proof).
Challenge phase: A selects a non-trivial challenge I∗
Enc and a
keyword w∗
and gives them to the challenger. The challenger
selects I∗
KeyGen such that F(I∗
KeyGen, I∗
Enc) = 1, generates
credential sk∗
with I∗
KeyGen and returns to A a search token
tk∗
by running tk∗
← TokenGen(sk, w∗
).
Guess: A outputs (rslt∗
, proof∗
) to the challenger. We say A
wins the game if 1 ← Verify(sk∗
, w∗
, tk∗
, rslt∗
, proof∗
) and
rslt∗
= rslt, where (rslt, proof) is produced by the challenger
by running SearchIndex(Au, Index, tk∗
).
Definition 5: A VABKS scheme is verifiable if the advantage
that any A wins the verifiability game is negligible in security
parameter ℓ.
C. Construction
A trivial solution for achieving verifiability is that a data
user downloads the keyword ciphertexts and conduct the
search operations locally. This solution incurs prohibitive
communication and computational overhead. As highlighted
in Figure 2, we instead let a data user outsource the key-
word search operation to the cloud, and then verify that the
cloud faithfully performed the keyword search operation. More
specifically, the data owner uses the signatures and bloom
filters as follows:
• A keyword signature is generated for each keyword
ciphertext and its associated data ciphertexts. It is used

Y
X
F2
F1
F3
W
V
Keywordsencrypted with
access control policy1
1: Encrypt keywords and data files, and generatekeyword
signature (e.g., !X, !y) for each keyword ciphertext and its
associateddataciphertexts.
2: Generatea bloom filterfor each keywordgroup, encrypt a
randomnumber thatis used for masking the bloomfilter, and
generatea bloom filtersignature (e.g., !BF1) for the masked bloom
filter(e.g. BF’1) and the randomnumber ciphertext(e.g., CphBF1 ).
3: For each keyword group, generate a local signature (e.g., !1)
for all keyword ciphertexts.
4: Generatea global signature (e.g., !) for all randomnumber
ciphertexts.
Keywordsencrypted with
access control policy2
!1
!2
!
CphF2
CphF1
CphF3
CphY
CphX
CphW
CphV
!X
!y
!v
!w
BF’1
CphBF1
!BF1
BF’2
CphBF2
!BF2
Fig. 2. Basic idea for achieving verifiability, where data files F1, F2, F3 were encrypted to cphF1
, cphF2
, cphF3
, keywords X, Y were encrypted to
cphX , cphY with access control policy 1, and keywords V, W were encrypted to cphV , cphW with access control policy 2. Given a search token tk, for
keyword group i, the cloud provides (σw, cphBFi
) as the proof when it finds keyword ciphertext cphw that matches tk, and (cphBFi
, BF′
i, σBFi
) otherwise.
for preventing the cloud from returning incorrect data
ciphertexts as the search result.
• For each keyword group, one bloom filter is built from
its keywords. This allows a data user to check that the
searched keyword was indeed not in the keyword group
when the cloud returns a null search result, without
downloading all keyword ciphertexts from the cloud. A
random number is selected and encrypted with the same
access control policy as keywords. The random number
masks the bloom filter for preserving keyword privacy. A
bloom filter signature is generated for the masked bloom
filter and the random number ciphertext for assuring their
integrity.
• A global signature is obtained by signing random number
ciphertexts of all groups. It allows a data user to verify
the integrity of the random number ciphertexts.
• A local signature is generated for all keyword ciphertexts
within the same keyword group KGj. This signature
allows the user to validate the integrity of keyword
ciphertexts within the keyword group.
Figure 3 describes the VABKS scheme, which uses a
signature scheme Sig = (KeyGen, Sign, Verify), a symmet-
ric encryption scheme SE = (KeyGen, Enc, Dec), an ABE
scheme ABE = (Setup, KeyGen, Enc, Dec), where the latter
two encryption schemes are used to encrypt data files. The
VABKS scheme is built on top of an ABKS scheme ABKS =
(Setup, KeyGen, Enc, TokenGen, Search), which encrypts the
keywords. Note that ABE and ABKS can be their ciphertext-
policy variant or their key-policy variant, but for the same
type. This leads to two variants of VABKS.
Note that in the Verify algorithm of Figure 3, when an
authorized data user verifies a null search result for keyword
group {cphw|w ∈ KGi}, where the user searches keyword
w′
, it can happen that 1 ← BFVerify({H′
1, . . . , H′
k}, BFi, w′
)
due to the false-positive of the Bloom filter. To validate the
search result in this case, the Verify algorithm has to download
{cphw|w ∈ KGi}, and checks the keyword ciphertexts one by
one. We stress that this does not incur significant communica-
tion cost on average because we can set the false-positive rate
as low as possible by choosing appropriate m and k (i.e., upon
one search request, the “wasted” bandwidth communication
and computational cost are proportional to this false-positive
rate). For example, in our experiment we set the false-positive
rate to be 4.5 × 10−9
.
D. Security Analysis
Security of the VABKS scheme can be proven as the
following theorems, whose proofs are deferred to Appendix
D.
Theorem 5: If ABE and SE are secure against the chosen-
plaintext attack, the VABKS scheme achieves the data secrecy.
Theorem 6: If ABE is secure against chosen-plaintext at-
tack, H is a secure pseudorandom generator and ABKS is
selectively secure against chosen keyword attack, the VABKS
scheme is selectively secure against chosen-keyword attack.
Theorem 7: If ABE is secure against chosen-plaintext attack,
H is a secure pseudorandom generator and ABKS achieves
keyword secrecy, the VABKS scheme achieves keyword se-
crecy.
Theorem 8: If Sig is a secure signature, the VABKS con-
struction achieves the verifiability.
V. PERFORMANCE EVALUATION
We evaluate the efficiency of the ABKS schemes in terms
of both asymptotic complexity and actual execution time,
and the efficiency of the VABKS scheme in terms of actual
execution time. We do not consider the asymptotic complexity
of VABKS because it uses multiple building-blocks (e.g.,
signing and ABE schemes) that can be instantiated with any
secure solutions. Asymptotic complexity is measured in terms
of four kinds of operations: H1 denotes the operation of
mapping a bit-string to an element of G, Pair denotes the
pairing operation, E denotes the exponentiation operation in G,
and ET denotes the exponentiation operation in GT . We ignore
multiplication and hash operations (other than H1) because
they are much more efficient than the above operations [38].
We implemented ABKS and VABKS in JAVA, while using
the Java Pairing Based Cryptography library (jPBC) [38].

Init(1ℓ
): Given security parameter ℓ, the attribute authority chooses k universal hash functions H′
1, . . . , H′
k, which are
used to construct a m-bit Bloom filter. Let H : {0, 1}ℓ
→ {0, 1}m
be a secure pseudorandom generator, SE be a secure
symmetric encryption scheme, ABE be a secure ABE scheme and ABKS be a secure ABKS scheme. This algorithm executes
(ABE.pm, ABE.mk) ← ABE.Setup(1ℓ
) and (ABKS.pm, ABKS.mk) ← ABKS.Setup(1ℓ
). It sets the public parameter as
pm = (ABE.pm, ABKS.pm, H′
1, . . . , H′
k) and mk = (ABE.mk, ABKS.mk).
KeyGen(mk, IKeyGen): The attribute authority runs ABE.sk ← ABE.KeyGen(ABE.mk, IKeyGen) and ABKS.sk ←
ABKS.KeyGen(ABKS.mk, IKeyGen), sets sk = (ABE.sk, ABKS.sk), and sends sk to a data owner/user over an authenticated
private channel.
BuildIndex({IEnc}l, {I′
Enc}n, D): The data owner runs (Sig.sk, Sig.pk) ← Sig.KeyGen(1ℓ
), keeps Sig.sk private and makes
Sig.pk public. Given D = (KG = {KG1, . . . , KGl}, MP = {MP(w)|w ∈ ∪l
i=1KGi}, FS = {F1, . . . , Fn}), the data owner
executes as follows:
1) Encrypt each data file with hybrid encryption: ∀Fj ∈ FS, generate ciphertext cphFj
= (cphskj
, cphSEj
) by running
SE.skj ← SE.KeyGen(1ℓ
), cphSEj
← SE.Enc(SE.skj, Fj), and cphskj
← ABE.Enc(I′
Encj, SE.skj).
2) Encrypt each keyword and generate keyword signature: Given KGi, 1 ≤ i ≤ l, for each w ∈ KGi,
run cphw ← ABKS.Enc(IEnci
, w), set MP(cphw) = {IDcphFj
|IDFj ∈ MP(w)}, and generate σw ←
Sig.Sign(Sig.sk, cphw||string({cphFj
|IDcphFj
∈ MP(cphw)})), where IDFj and IDcphFj
are identifiers for identifying
data file Fj and data ciphertext cphFj
, respectively.
3) Generate a bloom filter, a bloom filter signature and a local signature for each group KGi: Let BFi ←
BFGen({H′
1, . . . , H′
k}, KGi), cphBFi
← ABE.Enc(IEnci
, M) for some randomly chosen M from the message
space of ABE, compute BF′
i = H(M) BFi and generate σBFi ← Sig.Sign(Sig.sk, BF′
||cphBFi
). Let σi ←
Sig.Sign(Sig.sk, string({cphw|w ∈ KGi})) .
4) Generate the global signature: Set σ = Sig.Sign(Sig.sk, cphBF1
|| . . . ||cphBFl
).
5) Let Au = (σ, σ1, . . . , σl, cphBF1
, . . . , cphBFl
, σBF1 , . . . , σBFl
, {σw|w ∈ ∪l
i=1KGi}), Index = ({cphw|w ∈
∪l
i=1KGi}, {MP(cphw)|w ∈ ∪l
i=1KGi}) and Dcph = ({cphFj
|Fj ∈ FS}).
TokenGen(sk, w): Given credentials sk, a data user generates search token tk ← ABKS.TokenGen(ABKS.sk, w).
SearchIndex(Au, Index, Dcph, tk): Let rslt be an empty set and proof = (σ) initially. The cloud enumerates i = {cphw|w ∈
KGi}, 1 ≤ i ≤ l, which are the keyword ciphertexts with respect to the same access control policy.
• For each cphw ∈ i, it runs γ ← ABKS.Search(cphw, tk). If γ = 0, it continues to process the next keyword ciphertext
in i; otherwise, it adds the tuple (cphw, {cphFj
|IDcphFj
∈ MP(cphw)}) to rslt and (σw, cphBFi
) to proof.
• If there exist no γ = 1 after processing all cphw in i, then its adds (BF′
i, cphBFi
, σBFi
) to proof.
Verify(sk, w, tk, proof, rslt): The data user verifies the search result from the cloud as follows:
1) Verify the integrity of the random number ciphertexts: Let γ = Sig.Verify(Sig.pk, σ, cphBF1
|| . . . ||cphBFl
). If γ = 0,
then return 0; otherwise, continue to execute the following.
2) For i = 1, . . . , l, it executes as follows to verify that the cloud indeed returned the correct result for each keyword
group i:
Case 1: If (cphw, {cphFj
|IDcphFj
∈ MP(cphw)}) ∈ rslt, meaning there exists the keyword ciphertext cphw, which
corresponds to the same access control policy as what is specified by cphBFi
, having the same keyword specified
by tk, then it runs γ ← ABKS.Search(cphw, tk) and γ′
← Sig.Verify(Sig.pk, σw, cphw||string({cphFj
|IDcphFj
∈
MP(cphw)})) to verify whether or not cphw matches tk and all the associated data ciphertexts are returned by the
cloud. If either γ = 0 or γ′
= 0, then return 0, otherwise, continue to i = i + 1.
Case 2: If (BF′
i, cphBFi
, σBFi
) ∈ proof meaning that there is no matching keyword ciphertext, then it continues to
verify the integrity of the masked Bloom filter by running γ′
← Sig.Verify(Sig.pk, σBFi
, BF′
i||cphBFi
). If γ′
= 0, return
0; otherwise, execute the following:
• If the data user is authorized, compute M ← ABE.Dec(ABE.sk, cphBFi
), BFi = H(M) BF′
i. Execute δ ←
BFVerify({H′
1, . . . , H′
k}, BFi, w) to check whether w or not is present in the keyword group as represented by
BFi.
– If δ = 0, meaning that w is not present in the keyword group as represented by BFi, then continue to i = i+1.
– If δ = 1, download i = {cphw|w ∈ KGi} and σi from the cloud, and run η ← Sig.Verify(Sig.pk, σi,
string({cphw|w ∈ KGi})). If η = 0, return 0; otherwise, run τ ← ABKS.Search(cphw, tk) by enumerating
cphw in cphw|w ∈ KGi}. If there exists some τ = 1 after processing all cphw (meaning that there exists some
cphw that matches tk), return 0; otherwise, continue to i = i + 1.
• If the data user is unauthorized, then it continues to i = i + 1 because cphBFi
cannot be decrypted.
Case 3: If none of the above two cases happens, return 0.
3) Return 1 if all tuples in the search result have been verified, and 0 otherwise.
Fig. 3. VABKS construction

In our implementation, the bilinear map is instantiated as
Type A pairing (ℓ = 512), which offers a level of security
equivalent to 1024-bit DLOG [38]. For both CP-VABKS and
KP-VABKS, we instantiated the symmetric encryption scheme
as AES-CBC, and the signature scheme with DSA provided
by JDK1.6. We instantiated ABKS, ABE as CP-ABKS, CP-
ABE [3] for CP-ABKS, and KP-ABKS, KP-ABE [2] for
KP-VABKS, respectively. Finally, we set the example access
control policy as “at1 AND . . . AND atN .”
A. Efficiency of ABKS
Asymptotic Complexity of the ABKS Schemes. Table I
describes the asymptotic complexities of the ABKS schemes.
We observe that in the CP-ABKS scheme, the complexity of
KeyGen is almost the same as that of Enc. In the KP-ABKS
scheme, KeyGen is more expensive than Enc. In both schemes,
the two Search algorithms incur almost the same cost.
complexity output size
KP-
KeyGen 3NE + NH1 2N|G|
Enc (S + 4)E + SH1 (S + 3)|G|
ABKS
TokenGen (2N + 2)E (2N + 2)|G|
Search (2S + 2)Pair + SET
CP-
KeyGen (2S + 2)E + SH1 (2S + 1)|G|
Enc (2N + 4)E + NH1 (2N + 3)|G|
ABKS
TokenGen (2S + 4)E (2S + 3)|G|
Search (2N + 3)Pair + NET
TABLE I
ASYMPTOTIC COMPLEXITIES OF CP-ABKS AND KP-ABKS, WHERE S IS
THE NUMBER OF A DATA USER’S ATTRIBUTES AND N IS THE NUMBER OF
ATTRIBUTES THAT ARE INVOLVED IN A DATA OWNER’S ACCESS CONTROL
POLICY (I.E., THE NUMBER OF LEAVES IN THE ACCESS TREE).
Actual Performance of the ABKS Schemes. To evaluate the
performance of the ABKS schemes, we ran the experiments
on a client machine with Linux OS, 2.93GHz Intel Core Duo
CPU (E7500), and 2GB RAM. We varied N, the number
of attributes that are involved in the example access control
policy, from 1 to 50 with step length 10. We ran each
experiment for 10 times to obtain the average execution time.
S/N
1 10 20 30 40 50
KP-
KeyGen 0.088 0.786 1.539 2.316 3.081 3.863
Enc 0.108 0.539 1.016 1.492 1.983 2.434
ABKS
TokenGen 0.073 0.331 0.627 0.917 1.211 1.504
Search 0.049 0.275 0.480 0.711 0.947 1.182
CP-
KeyGen 0.107 0.686 1.275 1.901 2.525 3.151
Enc 0.121 0.681 1.304 1.923 2.546 3.169
ABKS
TokenGen 0.088 0.349 0.673 0.932 1.228 1.513
Search 0.061 0.329 0.493 0.728 0.97 1.202
TABLE II
EXECUTION TIME (SECOND) OF THE ALGORITHMS IN THE KP -ABKS
AND CP -ABKS SCHEMES, WHERE N IS THE NUMBER OF ATTRIBUTES
INVOLVED IN THE EXAMPLE ACCESS CONTROL POLICY. THE NUMBER OF
DATA USER’S ATTRIBUTES IS ALSO SET TO N, NAMELY S = N IN THE
EXPERIMENTS.
Table II shows the execution time of the two ABKS
schemes. We observe that for both schemes, the keyword
encryption algorithm Enc (run by the data owner) is more
expensive than that of the keyword search algorithm Search
(run by the cloud) with the same N. However, the keyword
encryption algorithm is executed only once for each keyword,
whereas the keyword search algorithm will be performed as
many times as needed. Furthermore, we advocate that the data
users outsource the keyword search operations to the cloud
(i.e., taking advantage of the cloud’s computational resources).
B. Efficiency of VABKS with Real Data
To demonstrate the feasibility of VABKS in practice, we
evaluated it with real data, which consists of 2,019 distinct
keywords extracted from 670 PDF documents (papers) from
the ACM Digital Library with a total size of 778.1MB. We
set k = 28 and m = 10KB for Bloom filter so that
m
n = 10∗8∗1024
2019 ≈ 40 and the false-positive rate is around
4.5 × 10−9
. We vary the access control policy ranging from
1 to 50 attributes with step-length 10. In each experiment, we
encrypted all keywords with the same access control policy.
The algorithms run by the data owner and the data users
(i.e. BuildIndex, TokenGen and Verify) were executed on a
client machine with Linux OS, 2.93GHz Intel Core Duo CPU
(E7500), and 2GB RAM. The algorithm run by the cloud (i.e.,
SearchIndex) was executed on a server machine (a laptop) with
Windows 7, Intel i5 2.60GHz CPU, and 8GB RAM.
Figure 4(a) shows the execution time of BuildIndex that
was run by the data owner. We observe that with the same
attribute/policy complexity, CP-VABKS is more costly than
that of KP-VABKS when running algorithm BuildIndex. Figure
4(b) plots the execution time of the algorithms run by the data
user and the cloud. We simulated that algorithm SearchIndex
needs to conduct search operations over 1,010 keyword cipher-
texts to find the matched keyword ciphertext. We observe that
the execution time of TokenGen and Verify is really small
compared with keyword search algorithm SearchIndex. This
again confirms that the data user should outsource keyword
search operations to the cloud. Figure 4(c) plots the size
of index and auxiliary information, including 2,019 keyword
ciphertexts, bloom filters and signatures. We also see that CP-
VABKS consumes around two times more storage space than
KP-VABKS with the same attribute/policy complexity. These
discrepancies should serve as a factor when deciding whether
to use CP-VABKS or KP-VABKS in practice.
VI. CONCLUSION
We have introduced a novel cryptographic primitive called
verifiable attribute-based keyword search for secure cloud
computing over outsourced encrypted data. This primitive
allows a data owner to control the search of its outsourced
encrypted data according to an access control policy, while the
authorized data users can outsource the search operations to
the cloud and force the cloud to faithfully execute the search
(as a cheating cloud can be held accountable). Performance
evaluation shows that the new primitive is practical. Our study
focused on static data. As such, one interesting open problem
for future research is to accommodate dynamic data.
Acknowledgement. Zheng and Xu were supported in part by
the National Science Foundation under Grant No. 1111925.
Ateniese was supported by a Google Faculty Research Award,
an IBM Faculty Award, and the PRIN project TENACE.
REFERENCES
[1] A. Sahai and B. Waters, “Fuzzy identity-based encryption,” in Proc. of
EUROCRYPT, pp. 457–473, 2005.

1 10 20 30 40 50
0
1000
2000
3000
4000
5000
6000
7000
S(N)
Time(second)
CP−VAKBS.BuildIndex
KP−VABKS.BuildIndex
(a) BuildIndex
1 10 20 30 40 50
0
200
400
600
800
1000
S(N)
Time(second)
CP−VAKBS.SearchIndex
CP−VAKBS.TokenGen
CP−VAKBS.Verify
KP−VABKS.SearchIndex
KP−VABKS.TokenGen
KP−VABKS.Verify
(b) TokenGen, SearchIndex and Verify
1 10 20 30 40 50
0
5
10
15
20
25
30
S(N)
Indexsize(MB)
CP−VAKBS
KP−VABKS
(c) Size of index and auxiliary information
Fig. 4. Performance of the CP-VABKS and KP-VABKS schemes, where N is the number of attributes involved in the example access control policy. The
number of data user’s attributes is also set to N, namely S = N in the experiments.
[2] V. Goyal, O. Pandey, A. Sahai, and B. Waters, “Attribute-based encryp-
tion for fine-grained access control of encrypted data,” in Proc. of ACM
CCS, pp. 89–98, 2006.
[3] J. Bethencourt, A. Sahai, and B. Waters, “Ciphertext-policy attribute-
based encryption,” in Proc. of IEEE S&P, pp. 321–334, 2007.
[4] T. Okamoto and K. Takashima, “Fully secure functional encryption with
general relations from the decisional linear assumption,” in Proc. of
CRYPTO, pp. 191–208, 2010.
[5] A. B. Lewko and B. Waters, “New proof methods for attribute-based
encryption: Achieving full security through selective techniques,” in
Proc. of CRYPTO, pp. 180–198, 2012.
[6] M. Chase, “Multi-authority attribute based encryption,” in Proc. of TCC,
pp. 515–534, 2007.
[7] M. Chase and S. S. Chow, “Improving privacy and security in multi-
authority attribute-based encryption,” in Proc. of ACM CCS, pp. 121–
130, 2009.
[8] J. Camenisch, M. Kohlweiss, A. Rial, and C. Sheedy, “Blind and
anonymous identity-based encryption and authorised private searches
on public key encrypted data,” in Proc. of PKC, pp. 196–214, 2009.
[9] D. X. Song, D. Wagner, and A. Perrig, “Practical techniques for searches
on encrypted data,” in Proc. of IEEE S&P, pp. 44–, 2000.
[10] E.-J. Goh, “Secure indexes.” Cryptology ePrint Archive, Report
2003/216, 2003. http://eprint.iacr.org/2003/216/.
[11] Y.-C. Chang and M. Mitzenmacher, “Privacy preserving keyword
searches on remote encrypted data,” in Proc. of ACNS, pp. 442–455,
2005.
[12] R. Curtmola, J. Garay, S. Kamara, and R. Ostrovsky, “Searchable
symmetric encryption: improved definitions and efficient constructions,”
in Proc. of, pp. 79–88, 2006.
[13] M. Chase and S. Kamara, “Structured encryption and controlled disclo-
sure,” in Proc. of ASIACRYPT, pp. 577–594, 2010.
[14] K. Kurosawa and Y. Ohtaki, “Uc-secure searchable symmetric encryp-
tion,” in Proc. of FC, pp. 285–298, Springer Berlin / Heidelberg.
[15] S. Kamara and K. Lauter, “Cryptographic cloud storage,” in Proc. of
FC, pp. 136–149, 2010.
[16] S. Kamara, C. Papamanthou, and T. Roeder, “Cs2: A searchable cryp-
tographic cloud storage system.” Microsoft Technical Report, 2011.
http://research.microsoft.com/apps/pubs/?id=148632.
[17] S. Kamara, C. Papamanthou, and T. Roeder, “Dynamic searchable
symmetric encryption,” in Proc. of ACM CCS, pp. 965–976, 2012.
[18] Q. Chai and G. Gong, “Verifiable symmetric searchable encryption for
semi-honest-but-curious cloud servers,” in Proc. of ICC, pp. 917–922,
2012.
[19] D. Boneh, G. D. Crescenzo, R. Ostrovsky, and G. Persiano, “Public key
encryption with keyword search,” in Proc. of EUROCRYPT, pp. 506–
522, 2004.
[20] B. R. Waters, D. Balfanz, G. Durfee, and D. K. Smetters, “Building an
encrypted and searchable audit log,” in Proc. of NDSS, 2004.
[21] M. Bellare, A. Boldyreva, and A. O’Neill, “Deterministic and efficiently
searchable encryption,” in Proc. of CRYPTO, pp. 535–552, 2007.
[22] J. Baek, R. Safavi-Naini, and W. Susilo, “Public key encryption with
keyword search revisited,” in Proc. of ICCSA, pp. 1249–1259, 2008.
[23] P. Golle, J. Staddon, and B. Waters, “Secure conjunctive keyword search
over encrypted data,” in Proc. of ACNS, pp. 31–45, 2004.
[24] E. Shi, J. Bethencourt, H. T.-H. Chan, D. X. Song, and A. Perrig, “Multi-
dimensional range query over encrypted data,” in Proc. of IEEE S&P,
pp. 350–364, 2007.
[25] D. Boneh and B. Waters, “Conjunctive, subset, and range queries on
encrypted data,” in Proc. of TCC, pp. 535–554, 2007.
[26] M. Li, S. Yu, N. Cao, and W. Lou, “Authorized private keyword search
over encrypted data in cloud computing,” in Proc. of ICDCS, pp. 383–
392, 2011.
[27] F. Bao, R. H. Deng, X. Ding, and Y. Yang, “Private query on encrypted
data in multi-user settings,” in Proc. of ISPEC, pp. 71–85, 2008.
[28] T. Okamoto and K. Takashima, “Hierarchical predicate encryption for
inner-products,” in Proc. of ASIACRYPT, pp. 214–231, 2009.
[29] J. Katz, A. Sahai, and B. Waters, “Predicate encryption supporting
disjunctions, polynomial equations, and inner products,” in Proc. of
EUROCRYPT, pp. 146–162, 2008.
[30] S. Benabbas, R. Gennaro, and Y. Vahlis, “Verifiable delegation of
computation over large datasets,” in Proc. of CRYPTO, pp. 111–131,
2011.
[31] C. Papamanthou, E. Shi, and R. Tamassia, “Signatures of correct
computation.” Cryptology ePrint Archive, Report 2011/587, 2011. http:
//eprint.iacr.org/.
[32] D. Fiore and R. Gennaro, “Publicly verifiable delegation of large
polynomials and matrix computations, with applications,” in Proc. of
ACM CCS, pp. 501–512, 2012.
[33] D. Boneh, X. Boyen, and E.-J. Goh, “Hierarchical identity based
encryption with constant size ciphertext,” in Proc. of EUROCRYPT,
pp. 440–456, 2005.
[34] J. Katz and Y. Lindell, Introduction to Modern Cryptography. Chapman
and Hall/CRC Press, 2007.
[35] B. H. Bloom, “Space/time trade-offs in hash coding with allowable
errors,” Commun. ACM, vol. 13, pp. 422–426, July 1970.
[36] R. Canetti, S. Halevi, and J. Katz, “Chosen-ciphertext security from
identity-based encryption,” in EUROCRYPT, pp. 207–222, 2004.
[37] E. Shen, E. Shi, and B. Waters, “Predicate privacy in encryption
systems,” in Proc. of TCC, pp. 457–473, 2009.
[38] “The java pairing based cryptography library.
http://gas.dia.unisa.it/projects/jpbc/.”
APPENDIX A
PROOF OF THEOREM 1
Proof 1: We show that if there is a polynomial-time ad-
versary A that wins the SCKA game with advantage µ, then
there is a challenger algorithm that solves the DL problem
with advantage µ/2. Given a DL instance (g, h, f, fr1
, gr2
, Q),
where g, f, h, Q ← G and r1, r2 ← Zp, the challenger
simulates the SCKA game as follows.
Setup: The challenger sets ga
= h and gc
= f where a and
c are unknown, selects d ← Zp and computes gb
= fd
= gcd
by implicitly defining b = cd. Let H2 be an one-way hash
function and pm = (e, g, p, h, fd
, f) and mk = (d).
A selects an attribute set Atts∗
and gives it to the challenger.
The random oracle OH1 (atj) is defined as follows:

• If atj has not been queried before,
– if atj ∈ Atts∗
, select βj ← Zp, add (atj, αj = 0, βj)
to OH1 , and return gβj
;
– otherwise, select αj, βj ← Zp, add (atj, αj, βj) to
OH1 , and return fαj
gβj
.
• If atj has been queried before, retrieve (αj, βj) from OH1
and return fαj
gβj
.
Phase 1: A can adaptively query the following oracles for
polynomially-many times and the challenger keeps a keyword
list Lkw, which is empty initially.
OKeyGen(T): A gives an access tree T to the challenger. If
F(Atts∗
, T) = 1, then the challenger aborts; otherwise, the
challenger generates attributes as follows.
Define the following two procedures to determine the poly-
nomial for each node of T:
• PolySat(Tv, Atts∗
, λv): Given secret λv, this procedure
determines the polynomial for each node of Tv rooted at
v when F(Atts∗
, Tv) = 1. It works as follows: Suppose
the threshold value of node v is kv, it sets qv(0) = λv
and picks kv − 1 coefficients randomly to fix the poly-
nomial qv. For each child node v′
of v, recursively call
PolySat(Tv′ , Atts∗
, λv′ ) where λv′ = qv(Index(v′
)).
• PolyUnsat(Tv, Atts∗
, gλv
): Given element gλv
∈ G
where the secret λv is unknown, this procedure deter-
mines the polynomial for each node of Tv rooted at v
when F(Atts∗
, Tv) = 0 as follows. Suppose the threshold
value of the node v is kv. Let V be the empty set.
For each child node v′
of v, if F(Atts, Tv′ ) = 1, then
set V = V {v′
}. Because F(Atts, Tv) = 0, then
|V | < kv. For each node v′
∈ V , it selects λv′ ← Zp, and
sets qv(Index(v′
)) = λv′ . Finally it fixes the remaining
kv − |V | points of qv randomly to define qv and makes
gqv(0)
= gλv
. For each child node v′
of v,
– if F(Atts∗
, Tv′ ) = 1, then run
PolySat(Tv′ , Atts∗
, qv(Index(v′
)), where
qv(Index(v′
)) is known to the challenger;
– otherwise, call PolyUnsat(Tv′ , Atts∗
, gλv′
), where
gλv′
= gqv(Index(v′
)
is known to the challenger.
With the above two procedures, the challenger runs
PolyUnsat(T, Atts∗
, ga
), by implicitly defining qroot(0) = a.
Then for each v ∈ lvs(T), the challenger gets qv(0) if
att(v) ∈ Atts∗
, and gets gqv(0)
otherwise. Because cqv(0) is
the secret share of ac, due to the linear property, the challenger
generates credentials for each v ∈ lvs(T) as follows:
• If att(v) = atj for some atj ∈ Atts∗
: Select t ← Zp, set
Av = fqv(0)
gβjt
= gcqv(0)
H1(att(v))t
and Bv = gt
;
• If att(v) /∈ Atts∗
(assuming att(v) = atj): Select
t′
← Zp, set Av = (gqv(0)
)
−βj
αj (fαj
gβj
)t′
and Bv =
g
qv(0) −1
αj gt′
. Note that (Av, Bv) is a valid credential
because
Bv = g
qv(0) −1
αj gt′
= g
t′
−
qv(0)
αj
Av = g
qv(0)
−βj
αj (fαj
gβj
)t′
= fqv(0)
(fαj
gβj
)
−qv (0)
αj (fαj
gβj
)t′
= fqv(0)
(fαj
gβj
)
t′
− qv(0)
αj
= gcqv(0)
H1(att(v))
t′
− qv(0)
αj
by implicitly letting t = t′
− qv(0)
αj
. Note also that A
cannot construct Av and Bv without knowing αj, βj.
Eventually, the challenger returns sk = {(Av, Bv)|v ∈ lvs(T)}
to A.
OTokenGen(T, w): The challenger runs OKeyGen(T) to get sk =
(T, {Av, Bv|v ∈ lvs(T)}), computes tk ← TokenGen(sk, w),
and returns tk to A. If F(Atts, T) = 1, the challenger adds w
to the keyword List Lkw.
Challenge phase: A chooses two keywords w0 and w1 of
equal length, such that w0, w1 /∈ Lkw. The challenger outputs
cph∗
as:
• Select λ ← {0, 1}.
• For each atj ∈ Atts∗
, set Wj = (gr2
)βj
.
• Set W′
= fr1
, W = Q(fr1
)dH2(wλ)
, and W0 = gr2
.
• Set cph∗
= (Atts∗
, W′
, W, W0, {Wj|atj ∈ Atts∗
}) and
return cph∗
to A.
We note that if Q = hr1+r2
, then cph∗
is indeed a
legitimate ciphertext for keyword wλ. The reason is that
W′
= fr1
= gcr1
, W = Qfr1dH2(wλ)
= Qgr1cdH2(wλ)
=
ga(r1+r2)
gbr1H2(wλ)
, W0 = gr2
, and for atj ∈ Atts∗
, Wj =
(gr2
)βj
= H1(atj)r2
.
Phase 2: A continues to query the oracles as in Phase 1. The
only restriction is that (T, w0) and (T, w1) cannot be the input
to OTokenGen if F(Atts∗
, T) = 1.
Guess: Finally, A outputs a bit λ′
and gives it to the chal-
lenger. If λ′
= λ, then the challenger outputs Q = hr1+r2
;
otherwise, it outputs Q = hr1+r2
.
This completes the simulation. In the challenge phase, if
Q = hr1+r2
, then cph∗
is a valid ciphertext of wλ, so the
probability of A outputting λ = λ′
is 1
2 +µ. If Q is an element
randomly selected from G, then cph∗
is not a valid ciphertext
of wλ. The probability of A outputting λ = λ′
is 1
2 since
W is an random element in G. Therefore, the probability of
the challenger correctly guessing Q
?
= hr1+r2
with the DL
instance (g, h, f, fr1
, gr2
, Q) is 1
2 (1
2 + µ + 1
2 ) = 1
2 + µ
2 . That
is, the challenger solves the DL problem with advantage µ/2
if A wins the SCKA game with an advantage µ.
APPENDIX B
PROOF OF THEOREM 2
Proof 2: We construct a challenger that exploits the keyword
secrecy game as follows:
Setup: The challenger selects a, b, c ← Zp, f ← G. Let H2 be
an one-way hash function and pm = (e, g, ga
, gb
, gc
, f) and
mk = (a, b, c).

The random oracle OH1 (atj) is simulated as follows: If
atj has not been queried before, the challenger selects αj ←
Zp, adds (atj, αj) to OH1 , and returns gαj
; otherwise, the
challenger retrieves αj from OH1 and returns gαj
.
Phase 1: A can adaptively query the following oracles for
polynomially-many times.
OKeyGen(T): The challenger generates sk ← KeyGen(T, mk)
and returns sk to A. It adds T to the list LKeyGen, which is
initially empty.
OTokenGen(T, w): The challenger runs OKeyGen(T) to ob-
tain sk = (T, {Av, Bv|v ∈ lvs(T)}), computes tk ←
TokenGen(sk, w), and returns tk to A.
Challenge Phase: A selects an attribute set Atts∗
. The
challenger chooses an access control policy that is repre-
sented as T∗
such that F(Atts∗
, T∗
) = 1, computes sk∗
←
KeyGen(mk, T∗
). By taking as input Atts∗
and sk∗
, it selects
w∗
from keyword space uniformly at random, and computes
cph∗
and tk∗
with Enc and TokenGen. Atts∗
should satisfy
the requirement defined in the keyword secrecy game.
Guess: Finally, A outputs a keyword w′
and gives it to the
challenger. The challenger computes cph′
← Enc(Atts, w′
)
and if Search(tk∗
, cph′
) = 1, then A wins the game.
This finishes the simulation. Suppose A has already at-
tempted q distinct keywords before outputting w′
, we can see
that the probability of A winning the keyword secrecy game
is at most 1
|M|−q +ǫ. This is because the size of the remaining
keyword space is |M|−q, and as the H2 is an one way secure
hash function, meaning deriving w∗
from H2(w∗
) is at most a
negligible probability ǫ. Therefore, given q distinct keywords
A has attempted, the probability of A winning the keyword
secrecy game is at most 1
|M|−q +ǫ. Thus, our scheme achieves
keyword secrecy as in Definition 3.
APPENDIX C
PROOF OF THEOREM 3 ON CP-ABKS
Proof 3: We show that the CP-ABE scheme is selectively
secure against chosen-keyword attack in the generic bilinear
group model, where H1 is modeled as a random oracle and
H2 is a one-way hash function.
In the SCKA game, A attempts to distinguish
ga(r1+r2)gbr1H2(w0)
from ga(r1+r2)gbr1H2(w1)
. Given θ ← Zp,
the probability of distinguishing ga(r1+r2)
gbr1H2(w0)
from gθ
is equal to that of distinguishing gθ
from ga(r1+r2)
gbr1H2(w1)
.
Therefore, if A has advantage ǫ in breaking the SCKA game,
then it has advantage ǫ/2 in distinguishing ga(r1+r2)
gbr1H2(w0)
from gθ
. Thus, let us consider a modified game where A can
distinguish ga(r1+r2)
from gθ
. The modified SCKA game is
described as follows:
Setup: The challenger chooses a, b, c ← Zp and sends public
parameters (e, g, p, ga
, gb
, gc
) to A. A chooses an access tree
T∗
, which is sent to the challenger.
H1(atj) is simulated as follows: If atj has not been queried
before, the challenger chooses αj ← Zp, adds (atj, αj) to
OH1 and returns gαj
; otherwise the challenger returns gαj
by
retrieving αj from OH1 .
Phase 1: A can query OKeyGen and OTokenGen as follows:
a r
(t)
j s(ac + r(t))/b cr1
b r(t) + αj r
(t)
j s(r
(t)
j ) qv(0)
c (ac + r(t))/b s(r(t) + αj r
(t)
j ) αj qv(0)
αj cs s(a + bH2(w)) br2
TABLE III
POSSIBLE TERMS FOR QUERYING GROUP ORACLE GT
OKeyGen(Atts): The challenger selects r(t)
← Zp and com-
putes A = g(ac+r(t)
)/b
. For each attribute atj ∈ Atts, the
challenger chooses r
(t)
j ← Zp, computes Aj = gr(t)
gαj r
(t)
j
and Bj = gr
(t)
j , and returns (Atts, A, {(Aj, Bj)|atj ∈ Atts}).
OTokenGen(Atts, w): The challenger queries OKeyGen(Atts) to
get sk = (Atts, A, {(Aj, Bj)|atj ∈ Atts}) and returns tk =
(Atts, tok1, tok2, tok3, {(A′
j, B′
j)|atj ∈ Atts}) where tok1 =
(ga
gbH2(w)
)s
, tok2 = gcs
, tok3 = As
, A′
j = As
j and B′
j = Bs
j
by selecting s ← Zp. If F(Atts, T∗
) = 1, the challenger adds
w to the keyword List Lkw.
Challenge phase: Given two keywords w0, w1 of equal length
where w0, w1 /∈ Lkw, the challenger chooses r1, r2 ← Zp,
and computes secret shares of r2 for each leaves in T∗
. The
challenger selects λ ← {0, 1}. If λ = 0, it outputs
W = gcr1
, W0 = gθ
, W′
= gbr2
,
{(Wv = gqv(0)
, Dv = gαjqv(0)
)|v ∈ lvs(T∗
), att(v) = atj}
by selecting θ ∈ Zp; otherwise it outputs
W = gcr1
, W0 = ga(r1+r2)
, W′
= gbr2
,
{(Wv = gqv(0)
, Dv = gαjqv(0)
)|v ∈ lvs(T∗
), att(v) = atj}.
Phase 2: This is the same as in the SCKA game.
We can see that if A can construct e(g, g)δa(r1+r2)
for
some gδ
that can be composed from the oracle outputs he
has already queried, then A can use it to distinguish gθ
from
ga(r1+r2)
. Therefore, we need to show that A can construct
e(g, g)δa(r1+r2)
for some gδ
with a negligible probability. That
is, A cannot gain non-negligible advantage in the SCKA game.
In the generic group model, ψ0 and ψ1 are random injective
maps from Zp into a set of p3
elements. Then the probability of
A guessing an element in the image of ψ0 and ψ1 is negligible.
Recall that G = {ψ0(x)|x ∈ Zp} and GT = {ψ1(x)|x ∈
Zp}. Hence, let us consider the probability of A constructing
e(g, g)δa(r1+r2)
for some δ ∈ Zp from the oracle outputs he
has queried.
We list all terms that can be queried to the group oracle GT
in Table III. Let us consider how to construct e(g, g)δa(r1+r2)
for some δ. Because r1 only appears in the term cr1, δ
should contain c in order to construct e(g, g)δa(r1+r2)
. That
is, let δ = δ′
c for some δ′
and A wishes to construct
e(g, g)δ′
ac(r1+r2)
. Therefore, A needs to construct δ′
acr2,
which will use terms br2 and (ac+r(t)
)/b . Because (br2)(ac+
r(t)
)/b = acr2 +r(t)
r2, A needs to cancel r(t)
r2, which needs
to use the terms αj, r(t)
+ αjr
(t)
j , qv(0) and αjqv(0) because
qv(0) is the secret share of r2 according to T∗
. However, it is
impossible to construct r(t)
r2 with these terms because r(t)
r2

only can be reconstructed if the attributes corresponding to
r
(t)
j of r(t)
+ αjr
(t)
j satisfies the access tree T∗
.
Therefore, we can conclude that A gains a negligible
advantage in the modified game, which means that A gains a
negligible advantage in the SCKA game. This completes the
proof.
APPENDIX D
PROOFS OF THEOREM 5, THEOREM 6, THEOREM 7 AND
THEOREM 8 ON VABKS
A. Proof of Theorem 5 on VABKS
Proof 4: We show that if there exists a polynomial-time
algorithm A breaks VABKS’s data secrecy with the advantage
ρ, then we can break either CPA security for ABE or CPA
security for SE with the advantage ρ
n2 where n is the number
of data files to be encrypted.
The challenger proceeds the conventional CPA security
game with A. In the challenge phase, suppose A presents
two data collections D0 = (KG, MP, FS0 = {F01, . . . , F0n}),
D1 = (KG, MP, FS1 = {F11, . . . , F1n}), {IEnc}l and {I′
Enc}n.
The challenge selects λ ← {0, 1} and encrypts FSλ with the
ABE and {I′
Enc}n.
Now let us consider the advantage of A correctly guess-
ing λ. As we know, given two messages, the advantage of
distinguishing which message was encrypted by the hybrid
encryption of ABE and SE is equal. Therefore, given two sets
of data files FS0 and FS1, if the advantage of distinguishing
which data set was encrypted is ρ, then the advantage of
distinguishing which data file was encrypted is ρ
n2 by selecting
one data file from FS0 and one from FS1.
Therefore, we can see that if A breaks VABKS’s data se-
crecy of with a non-negligible advantage ρ, then the advantage
of breaking CPA security for ABE or CPA security for SE is ρ
n2
–a non-negligible probability, which contracts the assumption
that ABE is CPA-secure and SE is CPA-secure.
B. Proof of Theorem 6 on VABKS
Proof 5: We show that if there exists a polynomial-time
algorithm A breaks the selective security against chosen-
keyword attack of ABKS with the advantage ρ, then we can
break the selective security against chosen-keyword attack
game of ABKS with the advantage of ρ
l2 , given that ABE is
CPA-secure and H is a secure pseudorandom generator.
The challenger proceeds selective security against chosen-
keyword attack game with A. In the challenge phase, sup-
pose A presents two data collections D0 = (KG0 =
{KG01, . . . , KG0l}, MP, FS), and {I′
Enc}n. The challenge se-
lects λ ← {0, 1} and encrypts KG with ABKS, and generates
BF′
i, cphBFi
and σi for each keyword group.
Since ABE is CPA-secure and H is a secure pseudorandom
generator, the probability of A inferring λ via BF′
i, cphBFi
is
negligible. Then let us consider the advantage of A correctly
guessing λ from keyword ciphertexts. As we know, given two
keywords, the advantage of distinguishing which keyword was
encrypted by ABKS is equal. Therefore, given two keyword
sets KG0 and KG1, if the advantage of distinguishing which
keyword set was encrypted is ρ, then the advantage of distin-
guishing which keyword was encrypted is ρ
l2 by selecting one
keyword from KG0 and one from KG1.
Therefore, we can see that if A breaks VABKS’s selective
security against chosen-keyword attack with a non-negligible
advantage ρ, then the advantage of breaking ABKS’s selective
security against chosen-keyword attack is ρ
l2 –a non-negligible
probability, which contracts the assumption that ABKS achieve
selective security against chosen-keyword attack, given that
ABE is CPA-secure and H is a secure pseudorandom genera-
tor.
C. Proof of Theorem 7 on VABKS
Proof 6: We show that if there exists polynomial time
algorithm A breaking VABKS’skeyword secrecy, then it breaks
the assumption that ABKS achieves keyword secrecy.
Suppose A presents a data collection D = (KG =
{KG1, . . . , KGl}, MP, FS), {IEnc}l and {I′
Enc}n. The chal-
lenger simulates the keyword secrecy game as in B, where
the keyword space consists of keywords specified by FS.
We can see that the probability of A inferring the keyword
from a search token and corresponding keyword ciphertext is
equal to that of ABKS. Therefore, if in VABKSA guesses the
keyword from the search token and corresponding keyword
ciphertext with the probability more than 1
|M|−q + ǫ after
guessing q distinct keywords, then the probability of guessing
the keyword from the search token and keyword ciphertext
in ABKS is more than 1
|M|−q + ǫ after guessing q distinct
keywords, which contracts the assumption that ABKS achieves
keyword secrecy.
D. Proof of Theorem 8 on VABKS
Proof 7: We show that under the assumptions that Sig is
unforgeable, any polynomial-time adversary A presents an
incorrect search result and succeeds in the verification with
negligible probability.
The challenger proceeds the verifiability game, where
A provides the keyword-based data D = (KG =
{KG1, . . . , KGl}, MP = {MP(w)|w ∈ ∪l
i=1KGi}, FS =
{F1, . . . , Fn}), {IEnc}l and {I′
Enc}n. The challenger runs
(Au, Index, Dcph) ← BuildIndex({IEnc}l, {I′
Enc}n, D), and
gives (Au, Index, Dcph) to A.
In the challenge phase, with w∗
and I∗
Enc from A, the
challenger selects I∗
KeyGen such that F(I∗
KeyGen, I∗
Enc) = 1
where I∗
Enc is selected by A, generates credential sk∗
with
I∗
KeyGen and returns to A a search token tk∗
by running
tk∗
← TokenGen(sk, w∗
). A returns (rslt∗
, proof∗
) to the
challenger.
Suppose that (rslt∗
, proof∗
) succeeds in the verification.
That is, 1 ← Verify(sk∗
, w∗
, tk∗
, rslt∗
, proof∗
). Let us consider
the probability of A cheating with incorrect search result.
First, we claim that the global signature σ and random
keyword ciphertexts cphBF1
, . . . , cphBFl
are included in proof∗
without being manipulated; otherwise we can break the un-
forgeability of Sig.

Second, let us consider the search result within each group
with respect to access control policies, i.e. i = 1, . . . , l:
• If there exists no keyword ciphertext matched the search
token tk∗
, then we claim that A cannot cheat the chal-
lenger with some keyword ciphertext and data ciphertexts
in order to make VABKS.Verify output 1. The reason
is that A cannot forge a keyword signature σw for the
keyword ciphertext and data ciphertexts; otherwise, we
can break the unforgeability of Sig.
• If there exists a keyword ciphertext matched the search to-
ken tk∗
, then we claim that A cannot cheat the challenger
with a null search result in order to make VABKS.Verify
output 1. Suppose A returns a null result and the proof
(BF′
i, cphBFi
, σBFi
). Since BF′
i cannot be manipulated due
to σBFi
, the unmasked bloom ﬁlter indicates that w∗
is
a member within the group. The challenger downloads
cphw1
, . . . , cphw|KGi| and σi without being manipulated;
otherwise we break the Sig’s unforgeability. Then the
challenger can conduct the search operation with each
keyword ciphertext, and VABKS.Verify will output 0.
That is, if there exists keyword ciphertext matched the
search token, A returns a null result, then it cannot make
VABKS.Verify output 1.
To sum up, in order to make VABKS.Verify output 1, A has
to faithfully execute search operations and return the search
result honestly; otherwise, we will break Sig’s unforgeability.

Ieeepro techno solutions ieee java project - vabks verifiable attribute-based keyword search over outsourced encrypted data

More Related Content

What's hot (19)

Viewers also liked (17)

Similar to Ieeepro techno solutions ieee java project - vabks verifiable attribute-based keyword search over outsourced encrypted data (20)

More from hemanthbbc (10)

Recently uploaded (20)

Ieeepro techno solutions ieee java project - vabks verifiable attribute-based keyword search over outsourced encrypted data