Attribute-Based Encryption for

Fine-Grained Access Control of Encrypted Data

Vipul Goyal∗ Omkant Pandey† Amit Sahai‡ Brent Waters §

Abstract
As more sensitive data is shared and stored by third-party sites on the Internet, there
will be a need to encrypt data stored at these sites. One drawback of encrypting data,
is that it can be selectively shared only at a coarse-grained level (i.e., giving another
party your private key). We develop a new cryptosystem for fine-grained sharing of
encrypted data that we call Key-Policy Attribute-Based Encryption (KP-ABE). In
our cryptosystem, ciphertexts are labeled with sets of attributes and private keys are
associated with access structures that control which ciphertexts a user is able to decrypt.
We demonstrate the applicability of our construction to sharing of audit-log information
and broadcast encryption. Our construction supports delegation of private keys which
subsumes Hierarchical Identity-Based Encryption (HIBE).

1 Introduction
There is a trend for sensitive user data to be stored by third parties on the Internet. For
example, personal email, data, and personal preferences are stored on web portal sites such
as Google and Yahoo. The attack correlation center, dshield.org, presents aggregated
views of attacks on the Internet, but stores intrusion reports individually submitted by
users. Given the variety, amount, and importance of information stored at these sites,
there is cause for concern that personal data will be compromised. This worry is escalated
by the surge in recent attacks and legal pressure faced by such services.
One method for alleviating some of these problems is to store data in encrypted form.
Thus, if the storage is compromised the amount of information loss will be limited. One
disadvantage of encrypting data is that it severely limits the ability of users to selectively
share their encrypted data at a fine-grained level. Suppose a particular user wants to grant
decryption access to a party to all of its Internet traffic logs for all entries on a particular
range of dates that had a source IP address from a particular subnet. The user either needs
∗
This research was supported in part by NSF ITR/Cybertrust grants 0456717 and 0627781.
†
This research was supported in part by NSF ITR/Cybertrust grants 0456717 and 0627781.
‡
This research was supported in part by an Alfred P. Sloan Foundation Research Fellowship, an Intel
equipment grant, and NSF ITR/Cybertrust grants 0205594, 0456717 and 0627781.
§
This research was supported in part by NSF, the US Army Research Office Grant No. W911NF-06-
1-0316, and the Department of Homeland Security (DHS) and the Department of Interior (DOI) under
Contract No. NBCHF040146. Any opinions, finding and conclusions or recommendations expressed in this
material are those of the author(s) and do not necessarily reflect the views of the funding agencies.

1
to act as an intermediary and decrypt all relevant entries for the party or must give the
party its private decryption key, and thus let it have access to all entries. Neither one of
these options is particularly appealing. An important setting where these issues give rise
to serious problems is audit logs (discussed in more detail in Section 7).
Sahai and Waters [34] made some initial steps to solving this problem by introducing
the concept of Attributed-Based Encryption (ABE). In an ABE system, a user’s keys and
ciphertexts are labeled with sets of descriptive attributes and a particular key can decrypt
a particular ciphertext only if there is a match between the attributes of the ciphertext and
the user’s key. The cryptosystem of Sahai and Waters allowed for decryption when at least
k attributes overlapped between a ciphertext and a private key. While this primitive was
shown to be useful for error-tolerant encryption with biometrics, the lack of expressibility
seems to limit its applicability to larger systems.

Our Contribution. We develop a much richer type of attribute-based encryption cryp-

tosystem and demonstrate its applications. In our system each ciphertext is labeled by
the encryptor with a set of descriptive attributes. Each private key is associated with an
access structure that specifies which type of ciphertexts the key can decrypt. We call such a
scheme a Key-Policy Attribute-Based Encryption (KP-ABE), since the access structure is
specified in the private key, while the ciphertexts are simply labeled with a set of descriptive
attributes.1
We note that this setting is reminiscent of secret sharing schemes (see, e.g., [4]). Using
known techniques one can build a secret-sharing scheme that specifies that a set of parties
must cooperate in order to reconstruct a secret. For example, one can specify a tree access
structure where the interior nodes consist of AND and OR gates and the leaves consist of
different parties. Any set of parties that satisfy the tree can reconstruct the secret.
In our construction each user’s key is associated with a tree-access structure where the
leaves are associated with attributes.2 A user is able to decrypt a ciphertext if the attributes
associated with a ciphertext satisfy the key’s access structure. The primary difference
between our setting and secret-sharing schemes is that while secret-sharing schemes allow
for cooperation between different parties, in our setting, this is expressly forbidden. For
instance, if Alice has the key associated with the access structure “X AND Y”, and Bob
has the key associated with the access structure “Y AND Z”, we would not want them to
be able to decrypt a ciphertext whose only attribute is Y by colluding. To do this, we adapt
and generalize the techniques introduced by [34] to deal with more complex settings. We
will show that this cryptosystem gives us a powerful tool for encryption with fine-grained
access control for applications such as sharing audit log information.
In addition, we provide a delegation mechanism for our construction. Roughly, this
allows any user that has a key for access structure X to derive a key for access structure
Y, if and only if Y is more restrictive than X. Somewhat surprisingly, we observe that our
construction with the delegation property subsumes Hierarchical Identity-Based Encryption
1
This contrasts with what we call Ciphertext-Policy Attribute-Based Encryption (CP-ABE), where an
access structure (i.e. policy) would be associated to each ciphertext, while a user’s private key would be
associated with a set of attributes. KP-ABE and CP-ABE systems are useful in different contexts.
2
In fact, we can extend our scheme to work for any access structure for which a Linear Secret Sharing
Scheme exists (see appendix A).

2
[24, 21] and its derivatives [1].

1.1 Organization
We begin with a discussion of related work in Section 2. Next, we give necessary back-
ground information and our definitions of security in Section 3. We then present our first
construction and a proof of security in Section 4 and later generalize it to work for any LSSS
realizable access structure in Appendix A. We give a construction for the large universe
case in Section 5. We then show how to add the delegation property in Section 6. We
follow with a discussion of how our system applies to audit logs in Section 7. We discuss
the application of our construction to broadcast encryption in Section 8. Finally, we discuss
some interesting extensions and open problems in Section 9.

2 Related Work
Fine-grained Access Control. Fine-grained access control systems facilitate granting
differential access rights to a set of users and allow flexibility in specifying the access rights
of individual users. Several techniques are known for implementing fine grained access
control.
Common to the existing techniques (see, e.g., [26, 20, 38, 28, 23, 29] and the references
therein) is the fact that they employ a trusted server that stores the data in clear. Access
control relies on software checks to ensure that a user can access a piece of data only if
he is authorized to do so. This situation is not particularly appealing from a security
standpoint. In the event of server compromise, for example, as a result of a software
vulnerability exploit, the potential for information theft is immense. Furthermore, there
is always a danger of “insider attacks” wherein a person having access to the server steals
and leaks the information, for example, for economic gains. Some techniques (see, e.g., [2])
create user hierarchies and require the users to share a common secret key if they are in
a common set in the hierarchy. The data is then classified according to the hierarchy and
encrypted under the public key of the set it is meant for. Clearly, such methods have several
limitations. If a third party must access the data for a set, a user of that set either needs
to act as an intermediary and decrypt all relevant entries for the party or must give the
party its private decryption key, and thus let it have access to all entries. In many cases,
by using the user hierarchies it is not even possible to realize an access control equivalent
to monotone access trees.
In this paper, we introduce new techniques to implement fine grained access control. In
our techniques, the data is stored on the server in an encrypted form while different users
are still allowed to decrypt different pieces of data per the security policy. This effectively
eliminates the need to rely on the storage server for preventing unauthorized data access.

Secret-Sharing Schemes. Secret-sharing schemes (SSS) are used to divide a secret among
a number of parties. The information given to a party is called the share (of the secret) for
that party. Every SSS realizes some access structure that defines the sets of parties who
should be able to reconstruct the secret by using their shares.

3
 Shamir [35] and Blakley [7] were the first to propose a construction for secret-sharing
schemes where the access structure is a threshold gate. That is, if any t or more parties come
together, they can reconstruct the secret by using their shares; however, any lesser number
of parties do not get any information about the secret. Benaloh [6] extended Shamir’s idea
to realize any access structure that can be represented as a tree consisting of threshold
gates. Other notable secret-sharing schemes are [25, 15].
In SSS, one can specify a tree-access structure where the interior nodes consist of AND
and OR gates and the leaves consist of different parties. Any set of parties that satisfy
the tree can come together and reconstruct the secret. Therefore in SSS, collusion among
different users (or parties) is not only allowed but required.
In our construction each user’s key is associated with a tree-access structure where the
leaves are associated with attributes. A user is able to decrypt a ciphertext if the attributes
associated with a ciphertext satisfy the key’s access structure. In our scheme, contrary to
SSS, users should be unable to collude in any meaningful way.

Identity-Based Encryption and Extensions. The concept of Attribute-Based Encryp-

tion was introduced by Sahai and Waters [34], who also presented a particular scheme that
they called Fuzzy Identity-Based Encryption (FIBE). The Fuzzy-IBE scheme builds upon
several ideas from Identity-Based Encryption [10, 36, 18]. In FIBE, an identity is viewed
as a set of attributes. FIBE allows for a private key for an identity, ω, to decrypt to a
ciphertext encrypted with an identity, ω 0 , if and only if the identities ω and ω 0 are close
to each other as measured by the “set overlap” distance metric. In other words, if the
message is encrypted with a set of attributes ω 0 , a private key for a set of attributes ω
enables decrypting that message, if and only if |ω ∩ ω 0 | ≥ d, where d is fixed during the
setup time. Thus, FIBE achieves error tolerance making it suitable for use with biomet-
ric identities. However, it has limited applicability to access control of data, our primary
motivation for this work. Since the main goal in FIBE is error tolerance, the only access
structure supported is a threshold gate whose threshold is fixed at the setup time.
We develop a much richer type of attribute-based encryption. The private keys of
different users might be associated with different access structures. Our constructions
support a wide variety of access structures (indeed, in its most general form, every LSSS
realizable access structure), including a tree of threshold gates.
Yao et. al. [19] show how an IBE system that encrypts to multiple hierarchical identities
in a collusion-resistant manner implies a forward secure Hierarchical IBE scheme. They
also note how their techniques for resisting collusion attacks are useful in attribute-based
encryption. However, the cost of their scheme in terms of computation, private key size,
and ciphertext size increases exponentially with the number of attributes. We also note
that there has been other work that applied IBE techniques to access control, but did not
address our central concern of resisting attacks from colluding users [37, 14].

3 Background
We first give formal definitions for the security of Key-Policy Attribute Based Encryp-
tion (KP-ABE). Then we give background information on bilinear maps and our crypto-
graphic assumption. Finally, we give some background on linear secret-sharing schemes

4
and monotone span programs.

3.1 Definitions
Definition 1 (Access Structure [4]) Let {P1 , P2 , . . . , Pn } be a set of parties. A collec-
tion A ⊆ 2{P1 ,P2 ,...,Pn } is monotone if ∀B, C : if B ∈ A and B ⊆ C then C ∈ A. An access
structure (respectively, monotone access structure) is a collection (respectively, monotone
collection) A of non-empty subsets of {P1 , P2 , . . . , Pn }, i.e., A ⊆ 2{P1 ,P2 ,...,Pn } \{∅}. The
sets in A are called the authorized sets, and the sets not in A are called the unauthorized
sets.

In our context, the role of the parties is taken by the attributes. Thus, the access
structure A will contain the authorized sets of attributes. We restrict our attention to
monotone access structures. However, it is also possible to (inefficiently) realize general
access structures using our techniques by having the not of an attribute as a separate at-
tribute altogether. Thus, the number of attributes in the system will be doubled. From now
on, unless stated otherwise, by an access structure we mean a monotone access structure.
An (Key-Policy) Attribute Based Encryption scheme consists of four algorithms.

Setup This is a randomized algorithm that takes no input other than the implicit security
parameter. It outputs the public parameters PK and a master key MK.

Encryption This is a randomized algorithm that takes as input a message m, a set of

attributes γ, and the public parameters PK. It outputs the ciphertext E.

Key Generation This is a randomized algorithm that takes as input – an access structure
A, the master key MK and the public parameters PK. It outputs a decryption key D.

Decryption This algorithm takes as input – the ciphertext E that was encrypted under
the set γ of attributes, the decryption key D for access control structure A and the public
parameters PK. It outputs the message M if γ ∈ A.
We now discuss the security of an ABE scheme. We define a selective-set model for
proving the security of the attribute based under chosen plaintext attack. This model can
be seen as analogous to the selective-ID model [16, 17, 8] used in identity-based encryption
(IBE) schemes [36, 10, 18].

Selective-Set Model for ABE

Init The adversary declares the set of attributes, γ, that he wishes to be challenged upon.
Setup The challenger runs the Setup algorithm of ABE and gives the public parameters
to the adversary.
Phase 1 The adversary is allowed to issue queries for private keys for many access
structures Aj , where γ ∈
/ Aj for all j.
Challenge The adversary submits two equal length messages M0 and M1 . The challenger
flips a random coin b, and encrypts Mb with γ. The ciphertext is passed to the adversary.

5
Phase 2 Phase 1 is repeated.
Guess The adversary outputs a guess b0 of b.

The advantage of an adversary A in this game is defined as Pr[b0 = b] − 21 .

We note that the model can easily be extended to handle chosen-ciphertext attacks by
allowing for decryption queries in Phase 1 and Phase 2.

Definition 2 An attribute-based encryption scheme is secure in the Selective-Set model

of security if all polynomial time adversaries have at most a negligible advantage in the
Selective-Set game.

3.2 Bilinear Maps

We present a few facts related to groups with efficiently computable bilinear maps.
Let G1 and G2 be two multiplicative cyclic groups of prime order p. Let g be a generator
of G1 and e be a bilinear map, e : G1 × G1 → G2 . The bilinear map e has the following
properties:
1. Bilinearity: for all u, v ∈ G1 and a, b ∈ Zp , we have e(ua , v b ) = e(u, v)ab .
2. Non-degeneracy: e(g, g) 6= 1.
We say that G1 is a bilinear group if the group operation in G1 and the bilinear map
e : G1 × G1 → G2 are both efficiently computable. Notice that the map e is symmetric
since e(g a , g b ) = e(g, g)ab = e(g b , g a ).

3.3 The Decisional Bilinear Diffie-Hellman (BDH) Assumption

Let a, b, c, z ∈ Zp be chosen at random and g be a generator of G1 . The decisional BDH
assumption [8, 34] is that no probabilistic polynomial-time algorithm B can distinguish the
tuple (A = g a , B = g b , C = g c , e(g, g)abc ) from the tuple (A = g a , B = g b , C = g c , e(g, g)z )
with more than a negligible advantage. The advantage of B is
¯ ¯
¯ ¯
¯Pr[B(A, B, C, e(g, g)abc ) = 0] − Pr[B(A, B, C, e(g, g)z )] = 0¯

where the probability is taken over the random choice of the generator g, the random choice
of a, b, c, z in Zp , and the random bits consumed by B.

3.4 LSSS and Monotone Span Programs

In a linear secret-sharing scheme [4], realizing an access structure A, a third party called
the dealer holds a secret y and distributes the shares of y to parties such that y can be
reconstructed by a linear combination of the shares of any authorized set. Further, an
unauthorized set has no information about the secret y.
There is a close relation between LSSS and a linear algebraic model of computation
called monotone span programs (MSP) [27]. It has been shown that the existence of an
efficient LSSS for some access structure is equivalent to the existence of a small monotone
span program for the characteristic function of that access structure [27, 4]. The following
definition of MSP is a slightly altered version of the one presented in [4].

6
Definition 3 (Monotone Span Program) Let K be a field and {x1 , . . . , xn } be a set
of variables. A monotone span program over K is labeled matrix M̂ (M, ρ) where M is a
matrix over K, and ρ is a labeling of the rows of M by literals from {x1 , . . . , xn } (every row
is labeled by one literal).
A monotone span program accepts or rejects an input by the following criterion. For
every input set γ of literals, define the submatrix Mγ of M consisting of those rows whose
labels are in γ, i.e., rows labeled by some xi such that xi ∈ γ. The monotone span program
M̂ accepts γ if and only if ~1 ∈ span(Mγ ), i.e., some linear combination of the rows of
Mγ gives the all-one vector ~1. The MSP M̂ computes a Boolean function fM if it accepts
exactly those inputs γ where fM (γ) = 1. The size of M̂ is the number of rows in M .
Again, since the role of parties will be assumed by attributes in our context, each row
of the matrix M will be labeled by an attribute.

4 Construction for Access Trees

In the access-tree construction, ciphertexts are labeled with a set of descriptive attributes.
Private keys are identified by a tree-access structure in which each interior node of the
tree is a threshold gate and the leaves are associated with attributes. (We note that this
setting is very expressive. For example, we can represent a tree with “AND” and “OR”
gates by using respectively 2 of 2 and 1 of 2 threshold gates.) A user will be able to decrypt
a ciphertext with a given key if and only if there is an assignment of attributes from the
ciphertexts to nodes of the tree such that the tree is satisfied.

4.1 Access Trees

Access tree T . Let T be a tree representing an access structure. Each non-leaf node of the
tree represents a threshold gate, described by its children and a threshold value. If numx
is the number of children of a node x and kx is its threshold value, then 0 < kx ≤ numx .
When kx = 1, the threshold gate is an OR gate and when kx = numx , it is an AND gate.
Each leaf node x of the tree is described by an attribute and a threshold value kx = 1.
To facilitate working with the access trees, we define a few functions. We denote the
parent of the node x in the tree by parent(x). The function att(x) is defined only if x is a
leaf node and denotes the attribute associated with the leaf node x in the tree. The access
tree T also defines an ordering between the children of every node, that is, the children
of a node are numbered from 1 to num. The function index(x) returns such a number
associated with the node x. Where the index values are uniquely assigned to nodes in the
access structure for a given key in an arbitrary manner.

Satisfying an access tree. Let T be an access tree with root r. Denote by Tx the subtree
of T rooted at the node x. Hence T is the same as Tr . If a set of attributes γ satisfies the
access tree Tx , we denote it as Tx (γ) = 1. We compute Tx (γ) recursively as follows. If x is a
non-leaf node, evaluate Tx0 (γ) for all children x0 of node x. Tx (γ) returns 1 if and only if at
least kx children return 1. If x is a leaf node, then Tx (γ) returns 1 if and only if att(x) ∈ γ.

7
4.2 Our Construction
Let G1 be a bilinear group of prime order p, and let g be a generator of G1 . In addition,
let e : G1 × G1 → G2 denote the bilinear map. A security parameter, κ, will determine the
size of the groups. We also define
Q the Lagrange coefficient ∆i,S for i ∈ Zp and a set, S,
of elements in Zp : ∆i,S (x) = j∈S,j6=i x−j
i−j . We will associate each attribute with a unique
element in Z∗p .
Our construction follows.

Setup Define the universe of attributes U = {1, 2, . . . , n}. Now, for each attribute i ∈ U,
choose a number ti uniformly at random from Zp . Finally, choose y uniformly at random
in Zp . The published public parameters PK are
T1 = g t1 , . . . , T|U| = g t|U | , Y = e(g, g)y .
The master key MK is:
t1 , . . . , t|U| , y .
Encryption (M, γ, PK) To encrypt a message M ∈ G2 under a set of attributes γ,
choose a random value s ∈ Zp and publish the ciphertext as:
0
E = (γ, E = M Y s , {Ei = Tis }i∈γ ) .
Key Generation (T , MK) The algorithm outputs a key that enables the user to decrypt
a message encrypted under a set of attributes γ if and only if T (γ) = 1. The algorithm
proceeds as follows. First choose a polynomial qx for each node x (including the leaves)
in the tree T . These polynomials are chosen in the following way in a top-down manner,
starting from the root node r.
For each node x in the tree, set the degree dx of the polynomial qx to be one less than
the threshold value kx of that node, that is, dx = kx − 1. Now, for the root node r, set
qr (0) = y and dr other points of the polynomial qr randomly to define it completely. For
any other node x, set qx (0) = qparent(x) (index(x)) and choose dx other points randomly to
completely define qx .
Once the polynomials have been decided, for each leaf node x, we give the following
secret value to the user:
qx (0)
Dx = g ti where i = att(x) .
The set of above secret values is the decryption key D.

Decryption (E, D) We specify our decryption procedure as a recursive algorithm . For

ease of exposition we present the simplest form of the decryption algorithm and discuss
potential performance improvements in the next subsection.
We first define a recursive algorithm DecryptNode(E, D, x) that takes as input the
ciphertext E = (γ, E 0 , {Ei }i∈γ ), the private key D (we assume the access tree T is embedded
in the private key), and a node x in the tree. It outputs a group element of G2 or ⊥.
Let i = att(x). If the node x is a leaf node then:
( qx (0)
e(Dx , Ei ) = e(g ti , g s·ti ) = e(g, g)s·qx (0) if i ∈ γ
DecryptNode(E, D, x) =
⊥ otherwise

8
 We now consider the recursive case when x is a non-leaf node. The algorithm DecryptNode(E, D, x)
then proceeds as follows: For all nodes z that are children of x, it calls DecryptNode(E, D, z)
and stores the output as Fz . Let Sx be an arbitrary kx -sized set of child nodes z such that
Fz 6= ⊥. If no such set exists then the node was not satisfied and the function returns ⊥.
Otherwise, we compute:
Y ∆i,S0 (0) i=index(z)
Fx = Fz x , where S 0 ={index(z):z∈S }
x x
z∈Sx
Y ∆ 0 (0)
= (e(g, g)s·qz (0) ) i,Sx

z∈Sx
Y ∆ 0 (0)
= (e(g, g)s·qparent(z) (index(z)) ) i,Sx (by construction)
z∈Sx
Y s·qx (i)·∆ 0 (0)
= e(g, g) i,Sx

z∈Sx

= e(g, g)s·qx (0) (using polynomial interpolation)

and return the result.
Now that we have defined our function DecryptNode, the decryption algorithm sim-
ply calls the function on the root of the tree. We observe that DecryptNode(E, D, r) =
e(g, g)ys = Y s if and only if the ciphertext satisfies the tree. Since, E 0 = M Y s the decryp-
tion algorithm simply divides out Y s and recovers the message M .

4.3 Efficiency
We now consider the efficiency of the scheme in terms of ciphertext size, private key size,
and computation time for decryption and encryption.
The ciphertext overhead will be approximately one group element in G1 for every ele-
ment in γ. That is the number of group elements will be equal to the number of descriptive
attributes in the ciphertext. Similarly, the encryption algorithm will need to perform one
exponentiation for each attribute in γ.
The public parameters in the system will be of size linear in the number of attributes
defined in the system. (We will later discuss systems where the set of legal attributes is not
dependent upon the public parameters and we give a so called Large-Universe construction
in Section 5.) User’s private keys will consist of a group element for every leaf in the key’s
corresponding access tree.
The decryption procedure is by far the hardest to define performance for. In our rudi-
mentary decryption algorithm the number of pairings to decrypt might always be as large
as the number of nodes in the tree. However, this method is extremely suboptimal and we
now discuss methods to improve upon it.
One important idea is for the decryption algorithm to do some type of exploration of the
access tree relative to the ciphertext attributes before it makes cryptographic computations.
At the very least the algorithm should first discover which nodes are not satisfied and not
bother performing cryptographic operations on them.
The following observation shows how to modify our decryption method to optimize the
efficiency. First we find out which leaf nodes we should use in order to minimize the number

9
of pairing computations as follows. For each node x, define a set Sx . If x is a leaf node,
then Sx = {x}. Otherwise, let k be the threshold value of x. From among the child nodes
of x, choose k nodes x1 , x2 , . . . , xk such that Sxi (for i = 1, 2, . . . , k) are first k sets of the
smallest size. Then for non-leaf node x, Sx = {x0 : x0 ∈ Sxi , i = 1, 2, . . . , k}. The set
Sr corresponding to the root node r denotes the set of leaf nodes that should be used in
order to minimize the number of pairing computations. Next, we notice that in the given
decryption algorithm, Lagrange coefficients (∆i,Sx0 ) from various levels get multiplied in the
exponent in a certain way in Zp . Thus, instead of exponentiating at each level, for each
leaf node x ∈ Sr , we can keep track of which Lagrange coefficients get multiplied with each
other. Using this we can compute the final Q exponent fx for feach leaf node x ∈ Sr by doing
multiplication in Zp . Now Fr is simply x∈Sr e(Dx , Eatt(x) ) . This reduces the number of
x

pairing computations and exponentiations to |Sr |. Thus, decryption is dominated by |Sr |

pairing computations.
The number of group elements that compose a user’s private key grows linearly with
the number of leaf nodes in the access tree. The number of group elements in a ciphertext
grows linearly with the size of the set we are encrypting under. Finally, the number of
group elements in the public parameters grows linearly with the number of attributes in
the defined universe. Later, we provide a construction for large universes where all elements
in Z∗p can be used as attributes, yet the size of public parameters only grows linearly in a
parameter n that we set to be the maximum possible size of γ.

4.4 Proof of Security

We prove that the security of our scheme in the attribute-based Selective-Set model reduces
to the hardness of the Decisional BDH assumption.

Theorem 1 If an adversary can break our scheme in the Attribute-based Selective-Set

model, then a simulator can be constructed to play the Decisional BDH game with a non-
negligible advantage.

Proof: Suppose there exists a polynomial-time adversary A, that can attack our scheme
in the Selective-Set model with advantage ². We build a simulator B that can play the
Decisional BDH game with advantage ²/2. The simulation proceeds as follows:
We first let the challenger set the groups G1 and G2 with an efficient bilinear map, e
and generator g. The challenger flips a fair binary coin µ, outside of B’s view. If µ = 0,
the challenger sets (A, B, C, Z) = (g a , g b , g c , e(g, g)abc ); otherwise it sets (A, B, C, Z) =
(g a , g b , g c , e(g, g)z ) for random a, b, c, z. We assume the universe, U is defined.

Init The simulator B runs A. A chooses the set of attributes γ it wishes to be challenged
upon.

Setup The simulator sets the parameter Y = e(A, B) = e(g, g)ab . For all i ∈ U, it sets Ti
as follows: if i ∈ γ, it chooses a random ri ∈ Zp and sets Ti = g ri (thus, ti = ri ); otherwise
it chooses a random βi ∈ Zp and sets Ti = g bβi = B βi (thus, ti = bβi ). It then gives the
public parameters to A.

10
Phase 1 A adaptively makes requests for the keys corresponding to any access structures
T such that the challenge set γ does not satisfy T . Suppose A makes a request for the
secret key for an access structure T where T (γ) = 0. To generate the secret key, B needs
to assign a polynomial Qx of degree dx for every node in the access tree T .
We first define the following two procedures: PolySat and PolyUnsat.
PolySat(Tx , γ, λx ) This procedure sets up the polynomials for the nodes of an access sub-
tree with satisfied root node, that is, Tx (γ) = 1. The procedure takes an access tree
Tx (with root node x) as input along with a set of attributes γ and an integer λx ∈ Zp .
It first sets up a polynomial qx of degree dx for the root node x. It sets qx (0) = λx and
then sets rest of the points randomly to completely fix qx . Now it sets polynomials for
each child node x0 of x by calling the procedure PolySat(Tx0 , γ, qx (index(x0 ))). Notice
that in this way, qx0 (0) = qx (index(x0 )) for each child node x0 of x.

PolyUnsat(Tx , γ, g λx ) This procedure sets up the polynomials for the nodes of an access
tree with unsatisfied root node, that is, Tx (γ) = 0. The procedure takes an access
tree Tx (with root node x) as input along with a set of attributes γ and an element
g λx ∈ G1 (where λx ∈ Zp ).
It first defines a polynomial qx of degree dx for the root node x such that qx (0) = λx .
Because Tx (γ) = 0, no more than dx children of x are satisfied. Let hx ≤ dx be
the number of satisfied children of x. For each satisfied child x0 of x, the procedure
chooses a random point λx0 ∈ Zp and sets qx (index(x0 )) = λx0 . It then fixes the
remaining dx − hx points of qx randomly to completely define qx . Now the algorithm
recursively defines polynomials for the rest of the nodes in the tree as follows. For
each child node x0 of x, the algorithm calls:

– PolySat(Tx0 , γ, qx (index(x0 ))), if x0 is a satisfied node. Notice that qx (index(x0 ))

is known in this case.
0
– PolyUnsat(Tx0 , γ, g qx (index(x )) ), if x0 is not a satisfied node. Notice that only
0
g qx (index(x )) can be obtained by interpolation as only g qx (0) is known in this case.

Notice that in this case also, qx0 (0) = qx (index(x0 )) for each child node x0 of x.
To give keys for access structure T , simulator first runs PolyUnsat(T , γ, A) to define a
polynomial qx for each node x of T . Notice that for each leaf node x of T , we know qx
completely if x is satisfied; if x is not satisfied, then at least g qx (0) is known (in some cases
qx might be known completely). Furthermore, qr (0) = a.
Simulator now defines the final polynomial Qx (·) = bqx (·) for each node x of T . Notice
that this sets y = Qr (0) = ab. The key corresponding to each leaf node is given using its
polynomial as follows. Let i = att(x).
 Q (0)
g xti = g bqxri(0) = B qxr(0)
i if att(x) ∈ γ
Dx = Qx (0) bqx (0) qx (0)
g ti = g bβi = g βi otherwise

Therefore, the simulator is able to construct a private key for the access structure T .
Furthermore, the distribution of the private key for T is identical to that in the original

11
scheme.

Challenge The adversary A, will submit two challenge messages m0 and m1 to the
simulator. The simulator flips a fair binary coin ν, and returns an encryption of mν . The
ciphertext is output as:
E = (γ, E 0 = mν Z, {Ei = C ri }i∈γ )
If µ = 0 then Z = e(g, g)abc . If we let s = c, then we have Y s = (e(g, g)ab )c = e(g, g)abc ,
and Ei = (g ri )c = C ri . Therefore, the ciphertext is a valid random encryption of message
mν .
Otherwise, if µ = 1, then Z = e(g, g)z . We then have E 0 = mν e(g, g)z . Since z is
random, E 0 will be a random element of G2 from the adversaries view and the message
contains no information about mν .

Phase 2 The simulator acts exactly as it did in Phase 1.

Guess A will submit a guess ν 0 of ν. If ν 0 = ν the simulator will output µ0 = 0 to

indicate that it was given a valid BDH-tuple otherwise it will output µ0 = 1 to indicate it
was given a random 4-tuple.
As shown in the construction the simulator’s generation of public parameters and private
keys is identical to that of the actual scheme.
In the case where µ = 1 the adversary gains no information about ν. Therefore, we
have Pr[ν 6= ν 0 |µ = 1] = 12 . Since the simulator guesses µ0 = 1 when ν 6= ν 0 , we have
Pr[µ0 = µ|µ = 1] = 21 .
If µ = 0 then the adversary sees an encryption of mν . The adversary’s advantage in
this situation is ² by definition. Therefore, we have Pr[ν = ν 0 |µ = 0] = 21 + ². Since the
simulator guesses µ0 = 0 when ν = ν 0 , we have Pr[µ0 = µ|µ = 0] = 12 + ².
The overall advantage of the simulator in the Decisional BDH game is 21 Pr[µ0 = µ|µ =
0] + 12 Pr[µ0 = µ|µ = 1] − 12 = 12 ( 12 + ²) + 12 21 − 12 = 12 ².

Chosen-Ciphertext Security. Our security definitions and proofs have been in the
chosen-plaintext model. Similar to [34], we notice that our construction can be extended
to the chosen-ciphertext model by applying the technique of using simulation-sound NIZK
proofs to achieve chosen-ciphertext security [33]. However, in Section 9 we describe how our
delegation mechanism can be used with the techniques of Cannetti, Halevi, and Katz [17]
to achieve a much more efficient CCA-2 system.

5 Large Universe Construction

In our previous constructions, the size of public parameters grows linearly with the number
of possible attributes in the universe. Combining the tricks presented in section 4 with those
in the large universe construction of Sahai and Waters [34], we construct another scheme
that uses all elements of Z∗p as attributes, yet the public parameters grow only linearly in

12
a parameter n which we fix as the maximum size of the set we can encrypt under.3
This construction not only reduces the size of the public parameters but also allows us
to apply a collision resistant hash function H : {0, 1}∗ → Z∗p and use arbitrary strings, that
were not necessarily considered during public key setup, as attributes. For example we can
add any verifiable attribute, such as “Lives in Beverly Hills”, to a user’s private key.

5.1 Description
Let G1 be a bilinear group of prime order p, and let g be a generator of G1 . Additionally,
let e : G1 × G1 → G2 denote the bilinear map. A security parameter, κ, will determine
the size of the groups. Also define the Lagrange coefficient ∆i,S for i ∈ Zp and a set, S, of
elements in Zp , exactly as before. The data will be encrypted under a set γ of n elements4
of Z∗p . Our construction follows.

Setup (n) Choose a random value y ∈ Zp and let g1 = g y . Now choose a random element
g2 of G1 .
Next, choose t1 , . . . , tn+1 uniformly at random from G1 . Let N be the set {1, 2, . . . , n +
1}. Define a function T , as:
n+1
Y ∆i,N (X)
n
T (X) = g2X ti .
i=1
Xn
Function T can be viewed as the function g2 g h(X) for some n degree polynomial h. The
public parameters PK are: g1 , g2 , t1 , . . . , tn+1 and the master key MK is: y.

Encryption (m, γ, PK) To encrypt a message m ∈ G2 under a set of attributes γ, choose

a random value s ∈ Zp and publish the ciphertext as:
0 00
E = (γ, E = me(g1 , g2 )s , E = g s , {Ei = T (i)s }i∈γ ).

Key Generation (T , MK, PK) The algorithm outputs a key which enables the user to
decrypt a message encrypted under a set of attributes γ, if and only if T (γ) = 1. The
algorithm proceeds as follows. First choose a polynomial qx for each non-leaf node x in the
tree T . These polynomials are chosen in the following way in a top down manner, starting
from the root node r.
For each node x in the tree, set the degree dx of the polynomial qx to be one less than
the threshold value kx of that node, that is, dx = kx − 1. Now for the root node r, set
qr (0) = y and dr other points of the polynomial qr randomly to define it completely. For
any other node x, set qx (0) = qparent(x) (index(x)) and choose dx other points randomly to
completely define qx .
3
If we are willing to accept random oracles [5], it is possible to overcome the size-limitation on γ by
replacing the function T (X) in our construction (see Setup) by a hash function (see [31] for details). This
also improves the efficiency of the system.
4
With some minor modifications, which we omit for simplicity, we can encrypt to all sets of size ≤ n.

13
 Once the polynomials have been decided, for each leaf node x, we give the following
secret values to the user:
q (0)
Dx = g2x · T (i)rx where i = att(x)
rx
Rx = g

where rx is chosen uniformly at random from Zp for each node x. The set of above secret
pairs is the decryption key D.

Decryption (E, D) As for the case of small universe, we first define a recursive algorithm
DecryptNode(E, D, x) that takes as input the ciphertext E = (γ, E 0 , E 00 , {Ei }i∈γ ), the
private key D (we assume the access tree T is embedded in the private key), and a node x
in the tree. It outputs a group element of G2 or ⊥ as follows.
Let i = att(x). If the node x is a leaf node then:

 e(Dx ,E 00 ) q (0)
e(g x ·T (i)rx ,g s )
q (0)
e(g x ,g s )·e(T (i)rx ,g s )
e(R ,E ) = 2e(grx ,T (i)s ) = 2 e(grx ,T (i)s ) = e(g, g2 )s·qx (0) if i ∈ γ
DecryptNode(E, D, x) = x i
⊥ otherwise

We now consider the recursive case when x is a non-leaf node. The algorithm DecryptNode(E, D, x)
then proceeds as follows: For all nodes z that are children of x, it calls DecryptNode(E, D, z)
and stores the output as Fz . Let Sx be an arbitrary kx -sized set of child nodes z such that
Fz 6= ⊥. If no such set exists then the node was not satisfied and the function returns ⊥.
Otherwise, we compute:
Y ∆ 0 (0)
i,Sx i=index(z)
Fx = Fz , where 0
Sx ={index(z):z∈Sx }
z∈Sx
Y ∆ 0 (0)
= (e(g, g2 )s·qz (0) ) i,Sx

z∈Sx
Y ∆ 0 (0)
= (e(g, g2 )s·qparent(z) (index(z)) ) i,Sx (by construction)
z∈Sx
Y s·qx (0)·∆ 0 (0)
= e(g, g2 ) i,Sx

z∈Sx

= e(g, g2 )s·qx (0) (using polynomial interpolation)

and return the result.

Now that we have defined our function DecryptNode, the decryption algorithm sim-
ply calls the function on the root of the tree. We observe that DecryptNode(E, D, r) =
e(g, g2 )ys = e(g1 , g2 )s if and only if the ciphertext satisfies the tree. Since, E 0 = me(g1 , g2 )s
the decryption algorithm simply divides out e(g1 , g2 )s and recovers the message m.
The security proof of this construction follows largely from the security proof in sec-
tion 4.4. Details are given in appendix B.
It is possible to extend the above construction to realize any LSSS realizable access
structure using techniques similar to those in appendix A. A brief outline of the the
construction follows.

14
 Algorithms Setup and Encryption remain exactly the same as for the above construction.
Key Generation takes an MSP and the master key as input. For each row i of the matrix,
~
it gives the following pair to the user: (Di = g2Mi ·~u · T (i)ri , Ri = g ri ); where ~u is chosen
such that ~1 · ~u = y (master key), and ri is chosen randomly. Decryption and the security
proof follow using the tricks of appendix A.

6 Delegation of Private Keys

In our large universe construction, individual users can generate new private keys using
their private keys, which can then be delegated to other users. A user which has a private
key corresponding to an access tree T can compute a new private key corresponding to
ANY access tree T 0 which is more restrictive than T (i.e., T 0 ⊆ T ). Thus, the users are
capable to acting as a local key authority which can generate and distribute private keys
to other users.
Computation of a new private key from an existing private key is done by applying a
set of basic operations on the existing key. These operations are aimed at step by step
conversion of the given private key for an access tree T to a private key for the targeted
access tree T 0 (given that T 0 ⊆ T ). In the following, a (t, n)-gate denotes a gate with
threshold t and number of children n. The operations are as follows.

1) Adding a new trivial gate to T

This operation involves adding a new node y above an existing node x. The new node y
represents a (1, 1) threshold gate which after adding becomes the parent of x. The former
parent of x (if x is not the root node), say z, becomes the parent of y.
Since the threshold of y is 1, we are required to associate a 0 degree polynomial qy
with it such that qx (0) = qy (index(x)) and qy (0) = qz (index(y)). The second condition
essentially fixes qy and the first one is automatically satisfied since z was the parent of x
earlier. Hence, no changes to the private key are required for this operation.

2) Manipulating an existing (t, n)-gate in T

This operation involves manipulating a threshold gate so as to make the access structure
more restrictive. The operation could be of the following three types.

2.1) Converting a (t, n)-gate to a (t + 1, n)-gate with (t + 1) ≤ n

Consider a node x representing a (t, n)-gate. Clearly, the polynomial qx has the degree
(t − 1) which has to be increased to t. Define a new polynomial qx0 as follows.

qx0 (X) = (X + 1)qx (X)

Now, we change the key such that qx0 becomes the new polynomial of the node x. This
is done as follows. For every child y of x, compute the constant Cx = index(y) + 1. For
every leaf node z in the subtree5 Ty , compute the new decryption key as

Dz0 = (Dz )Cx , Rz0 = (Rz )Cx

5
Recall that Ty denotes the subtree defined by y as its root.

15
The above results in the multiplication of all the polynomials in the subtree Ty with the
constant Cx . Hence, qy0 (0) = (index(y) + 1)qy (0) which is indeed a point on the new
polynomial qx0 . Note that since qx0 (0) = qx (0), no changes outside the subtree Tx are
required.
The above procedure effectively changes x from a (t, n)-gate to a (t + 1, n)-gate and
yields the corresponding new private key.

2.2) Converting a (t, n)-gate to a (t + 1, n + 1)-gate

This procedure involves adding a new subtree (with root say z) as a child of a node x
while increasing the degree of x by 1 at the same time. Let z be the v th child of x so that
index(z) = v. We shall change the polynomial qx to the following.
−1
qx0 (X) = (aX + 1)qx (X) where a =
v
As in the previous operation, for every (existing) child y of x, the polynomials in the
subtree Ty are multiplied with the appropriate constant Cx = a.index(y) + 1. This ensures
that qy0 (0) is indeed a point on qx0 . Further, set qz (0) = 0 (= qx0 (v)). Given qz (0), keys
can be created for the subtree Tz as in the original key generation algorithm. Hence, the
keys of the subtrees of all the children (old as well as new) of the node x have been made
consistent with the new polynomial qx0 , thus achieving our goal.

2.3) Converting a (t, n)-gate to a (t, n − 1)-gate with t ≤ (n − 1)

This operation involves deleting a child y of a node x. This can be easily achieved just
by deleting decryption keys corresponding to all the leaves of Ty from the original decryp-
tion key.

3) Re-randomizing the obtained key

Once we obtain a key for the desired access structure (by applying a set of operations
of type 1 and 2), we apply a final re-randomization step to make it independent of the
original key from which it was computed. Re-randomization of a node x with a (known)
constant Cx is done as follows. Choose a random polynomial px of degree6 dx such that
px (0) = Cx . Define the new polynomial qx0 as qx0 (X) = qx (X) + px (X). We have to change
the key such that qx0 becomes the new polynomial of node x. This is done by recursively
re-randomizing every child y of x with the constant Cy = px (index(y)). If y is a leaf node,
the new decryption key corresponding to y is computed as follows.

Dy0 = Dy .g2 Cy .T (i)ry , Ry0 = Ry .g ry

Where i = att(y) and ry is chosen randomly.

Now, re-randomization of the private key is done just by re-randomizing the root node
r with the constant Cr = 0. The key obtained is the final key ready to be distributed to
other users.

6
Recall that dx is the degree of the polynomial qx associated with the node x

16
In the following theorem, we prove that the above set of operations is complete. That is,
give a key for an access tree T , this set of operations is sufficient to compute a key for ANY
access tree T 0 which is more restrictive than T .

Theorem 2 (Completeness Theorem) The given set of operations is complete.

Proof: We can obtain a key for an access tree T 0 from a key for T using the following
general technique. Add a new trivial gate x (using operation 1) above the root node of T
so that the new gate becomes the root node. Now we apply operation 2.2 to x to convert
it from a (1, 1)-gate to a (2, 2)-gate. The new subtree added as a child of x corresponds to
the access tree T 0 . Finally, we re-randomize the key obtained (operation 3). This gives us
a key for the access structure (T AND T 0 ). However, since T 0 is more restrictive than T ,
this access structure is equivalent to T 0 itself. Hence, we have obtained a key for the access
structure T 0 . We point out that this is a general method for obtaining a more restrictive
access structure; in practice users will likely use the delegation tools described above in a
more refined manner to achieve shorter private key sizes and faster decryption times.

In the above setting, we may imagine an entity having multiple private keys (procured from
different entities). We note that it is possible to use these multiple keys (for different access
structures) to compute a key for the targeted access structure. Given n keys for the access
trees T1 , T2 , ... ,Tn , using an operation similar to operation 1, we can connect them to
obtain a single tree with an OR gate being the root node and T1 , T2 , ... ,Tn each being a
child subtree of that OR gate. Thus, we obtain a single key for the access structure T = (T1
OR T2 OR ... OR Tn ). This key can then be used to generate new private keys.

7 Audit Log Application

An important application of KP-ABE deals with secure forensic analysis: One of the most
important needs for electronic forensic analysis is an “audit log” containing a detailed ac-
count of all activity on the system or network to be protected. Such audit logs, however,
raise significant security concerns: a comprehensive audit log would become a prized target
for enemy capture. Merely encrypting the audit log is not sufficient, since then any party
who needs to legitimately access the audit log contents (for instance a forensic analyst)
would require the secret key – thereby giving this single analyst access to essentially all
secret information on the network. Such problematic security issues arise in nearly every
secure system, and particularly in large-scale networked systems such as the Global Infor-
mation Grid, where diverse secret, top secret, and highly classified information will need to
appear intermingled in distributed audit logs.
Our KP-ABE system provides an attractive solution to the audit log problem. Audit
log entries could be annotated with attributes such as, for instance, the name of the user,
the date and time of the user action, and the type of data modified or accessed by the
user action. Then, a forensic analyst charged with some investigation would be issued a
secret key associated with a particular “access structure” – which would correspond to the
key allowing for a particular kind of encrypted search; such a key, for example, would only
open audit log records whose attributes satisfied the condition that “the user name is Bob,

17
OR (the date is between October 4, 2005 and October 7, 2005 AND the data accessed
pertained to naval operations off the coast of North Korea)”. Our system would provide
the guarantee that even if multiple rogue analysts collude to try to extract unauthorized
information from the audit log, they will fail.
A more concrete example audit-log application of our ABE system would be to the
ArmyCERT program, which uses netflow logs [30]. Basically, an entry is created for every
flow (e.g. TCP connection), indexed by seven attributes: source IP address, destination
IP address, L3 protocol type, source port, destination port, ToS byte (DSCP), and input
logical interface (ifIndex). These aspects of every flow are in the clear, and the payload can
be encrypted using our ABE system with these fields as attributes.
Note that in our scheme, we would need to assume that the attributes associated with
audit log entries would be available to all analysts.7 This may present a problem in highly
secret environments where even attributes themselves would need to be kept hidden from
analysts. We leave the problem of constructing KP-ABE systems where attributes associ-
ated with ciphertexts remain secret as an important open problem.

8 Application to Broadcast Encryption: Targeted Broadcast

We describe a new broadcast scenario that we call targeted broadcast. Consider the following
setting.
• A broadcaster broadcasts a sequence of different items, each one labeled with a set of
attributes describing the item. For instance, a television broadcaster might broadcast
an episode of the show “24”, and label this item with attributes such as the name of
the program (“24”), the genre (“drama”), the season, the episode number, the year,
month, and date of original broadcast, the current year, month, and date, the name
of the director, and the name of the producing company.

• Each user is subscribed to a different “package”. The user package describes an

access policy, which along with the set of attributes describing any particular item
being broadcast, determine whether or not the user should be able to access the item.
For example, a television user may want to subscribe to a package that allows him
view episodes of “24” from either the current season or Season 3. This could be
encoded as policy as (“24” AND (“Season:5” OR “Season:3”)).
The essential idea of Targeted Broadcast is to enjoy the economies-of-scale offered by
a broadcast channel, while still being able to deliver programming targeted at the needs
or wishes of individual users. The growing acceptability of such a model can be seen by
the rising popularity of DVR systems such as TiVo, which allow users to easily record only
the programming they want in order to watch it later. In the case of television, taking the
approach that we envision here would allow for much more flexibility than just allowing
users to select what channels they like.
Our KP-ABE system naturally offers a targeted broadcast system. A new symmetric
key would be chosen and used to encrypt each item being broadcast, and then the KP-ABE
7
We observe that this does not mean that the attributes need be “public.” Our KP-ABE system’s
ciphertexts could be re-encrypted, with a key that corresponds to the general clearance level of all analysts.

18
system would be used to encrypt the symmetric key with the attributes associated with
the item being broadcast. The KP-ABE system would precisely allow the flexibility we
envision in issuing private keys for the unique needs of each user.
It is worth mentioning that handling such a situation with the best known broadcast
encryption schemes [11, 22] (which allow encrypting to an arbitrary subset of users) is quite
inefficient in comparison. The efficiency of such systems is dependent on the size of the
authorized user set or the number of users in the system [11, 22], and would also require
the broadcaster to refer to its database of user authorizations each time a different item
is to be encrypted for broadcast. In our scheme, the encryption of an item would depend
only on the properties of that item. The broadcaster could in principle even forget about
the levels of access granted to each user after preparing a private key for the user.

9 Discussion and Extensions

We discuss various extensions to our scheme and open problems.

Achieving CCA-Security and HIBE from Delegation We briefly outline how we

can achieve efficient CCA-2 security and realize the Hierarchical Identity-Based Encryption
by applying delegation techniques to the large universe construction.
To achieve CCA-2 security an encyrptor will chooses a set γ of attributes to encrypt
the message under and then generate a public/private key pair for a one time signature
scheme. We let VK denote the bitstring represenation of the public key and let γ 0 be the set
γ ∪ VK. The encryptor encrypts the ciphertext under the attributes γ 0 and then signs the
ciphertext with the private key and attaches the signature and the public key description.
Suppose a user has a key for access structure X wishes to decrypt. The user first checks that
the ciphertext is signed under VK and rejects the ciphertext otherwise. Then it creates an
new key for the access structure of “X AND CCA : VK”. By similar arguments to those in
Canetti, Halevi, and Katz [17] this gives chosen-ciphertex security. We can also use other
methods [12, 13] to achieve greater efficiency.
We can realize a HIBE by simply managing the the assignment of attributes in a careful
manner. For example, to encrypt to the hierarchical identity “edu:ucla” one can encrypt
to the set of attributes { “1-edu”, “2-ucla” }. Someone who has the top-level key for edu
will have a policy that requires the attribute “1-edu” to be present. To delegate a key
for “edu:ucla” it simply creates a policy for “1-edu”AND “2-ucla” using our delegation
techniques. We view the fact that a primitive HIBE follows so simply from our scheme as
an attestation to the power of these techniques.

Ciphertext-Policy Attribute-Based Encryption. In this work, we considered the

setting where ciphertexts are associated with sets of attributes, whereas user secret keys
are associated with policies. As we have discussed, this setting has a number of natural
applications. Another possibility is to have the reverse situation: user keys are associated
with sets of attributes, whereas ciphertexts are associated with policies. We call such
systems Ciphertext-Policy Attribute-Based Encryption (CP-ABE) systems. We note that
the construction of Sahai and Waters [34] was most naturally considered in this framework.

19
CP-ABE systems that allow for complex policies (like those considered here) would have
a number of applications. An important example is a kind of sophisticated Broadcast
Encryption, where users are described by (and therefore associated with) various attributes.
Then, one could create a ciphertext that can be opened only if the attributes of a user match
a policy. For instance, in a military setting, one could broadcast a message that is meant to
be read only by users who have a rank of Lieutenant or higher, and who were deployed in
South Korea in the year 2005. We leave constructing such a system as an important open
problem.

Searching on Encrypted Data. Our current constructions do not hide the set of at-
tributes under which the data is encrypted. However, if it were possible to hide the at-
tributes, then viewing attributes as keywords in such a system would lead to the first
general keyword-based search on encrypted data [9]. A search query could potentially be
any monotone boolean formula of any number of keywords. We leave the problem of hiding
the set of attributes as open.

References
[1] Michel Abdalla, Dario Catalano, Alexander W. Dent, John Malone-Lee, Gregory
Neven, and Nigel P. Smart. Identity-based encryption gone wild. In Michele Bugliesi,
Bart Preneel, Vladimiro Sassone, and Ingo Wegener, editors, ICALP (2), volume 4052
of Lecture Notes in Computer Science, pages 300–311. Springer, 2006.

[2] S.G. Akl and P.D. Taylor. Cryptographic Solution to a Multi Level Security Problem.
In Advances in Cryptology – CRYPTO, 1982.

[3] H. Anton and C. Rorres. Elementary Linear Algebra, 9th Edition. 2005.

[4] A. Beimel. Secure Schemes for Secret Sharing and Key Distribution. PhD thesis, Israel
Institute of Technology, Technion, Haifa, Israel, 1996.

[5] M. Bellare and P. Rogaway. Random oracles are practical: A paradigm for designing
efficient protocols. In ACM conference on Computer and Communications Security
(ACM CCS), pages 62–73, 1993.

[6] J. Benaloh and Leichter J. Generalized Secret Sharing and Monotone Functions. In
Advances in Cryptology – CRYPTO, volume 403 of LNCS, pages 27–36. Springer, 1988.

[7] G. R. Blakley. Safeguarding cryptographic keys. In National Computer Conference,

pages 313–317. American Federation of Information Processing Societies Proceedings,
1979.

[8] D. Boneh and X. Boyen. Efficient Selective-ID Secure Identity Based Encryption
Without Random Oracles. In Advances in Cryptology – Eurocrypt, volume 3027 of
LNCS, pages 223–238. Springer, 2004.

20
 [9] D. Boneh, G.D. Crescenzo, R. Ostrovsky, and G. Persiano. Public-Key Encryption
with Keyword Search. In Advances in Cryptology – Eurocrypt, volume 3027 of LNCS,
pages 506–522. Springer, 2004.

[10] D. Boneh and M. Franklin. Identity Based Encryption from the Weil Pairing. In
Advances in Cryptology – CRYPTO, volume 2139 of LNCS, pages 213–229. Springer,
2001.

[11] D. Boneh, C. Gentry, and B. Waters. Collusion Resistant Broadcast Encryption with
Short Ciphertexts and Private Keys. In Advances in Cryptology – CRYPTO, volume
3621 of LNCS, pages 258–275. Springer, 2005.

[12] Dan Boneh and Jonathan Katz. Improved efficiency for cca-secure cryptosystems built
using identity-based encryption. In CT-RSA, pages 87–103, 2005.

[13] Xavier Boyen, Qixiang Mei, and Brent Waters. Direct chosen ciphertext security from
identity-based techniques. In ACM Conference on Computer and Communications
Security, pages 320–329, 2005.

[14] Robert W. Bradshaw, Jason E. Holt, and Kent E. Seamons. Concealing complex poli-
cies with hidden credentials. In ACM Conference on Computer and Communications
Security, pages 146–157, 2004.

[15] E. F. Brickell. Some ideal secret sharing schemes. Journal of Combinatorial Mathe-
matics and Combinatorial Computing, 6:105–113, 1989.

[16] R. Canetti, S. Halevi, and J. Katz. A Forward-Secure Public-Key Encryption Scheme.

In Advances in Cryptology – Eurocrypt, volume 2656 of LNCS. Springer, 2003.

[17] R. Canetti, S. Halevi, and J. Katz. Chosen Ciphertext Security from Identity Based
Encryption. In Advances in Cryptology – Eurocrypt, volume 3027 of LNCS, pages
207–222. Springer, 2004.

[18] Clifford Cocks. An identity based encryption scheme based on quadratic residues. In
IMA Int. Conf., pages 360–363, 2001.

[19] Y. Dodis, N. Fazio, A. Lysyanskaya, and D.F. Yao. ID-Based Encryption for Complex
Hierarchies with Applications to Forward Security and Broadcast Encryption. In ACM
conference on Computer and Communications Security (ACM CCS), pages 354–363,
2004.

[20] Rita Gavriloaie, Wolfgang Nejdl, Daniel Olmedilla, Kent E. Seamons, and Marianne
Winslett. No registration needed: How to use declarative policies and negotiation to
access sensitive resources on the semantic web. In ESWS, pages 342–356, 2004.

[21] Craig Gentry and Alice Silverberg. Hierarchical id-based cryptography. In ASI-
ACRYPT, pages 548–566, 2002.

[22] D. Halevy and A. Shamir. The LSD Broadcast Encryption Scheme. In Advances in
Cryptology – CRYPTO, volume 2442 of LNCS, pages 47–60. Springer, 2002.

21
[23] Hugh Harney, Andrea Colgrove, and Patrick Drew McDaniel. Principles of policy in
secure groups. In NDSS, 2001.

[24] Jeremy Horwitz and Ben Lynn. Toward hierarchical identity-based encryption. In
Lars R. Knudsen, editor, EUROCRYPT, volume 2332 of Lecture Notes in Computer
Science, pages 466–481. Springer, 2002.

[25] M. Ito, A. Saito, and T. Nishizeki. Secret Sharing Scheme Realizing General Access
Structure. In IEEE Globecom. IEEE, 1987.

[26] Myong H. Kang, Joon S. Park, and Judith N. Froscher. Access control mechanisms
for inter-organizational workflow. In SACMAT ’01: Proceedings of the sixth ACM
symposium on Access control models and technologies, pages 66–74, New York, NY,
USA, 2001. ACM Press.

[27] M. Karchmer and A. Wigderson. On Span Programs. In The Eighth Annual Structure
in Complexity Theory, pages 102–111, 1993.

[28] Jiangtao Li, Ninghui Li, and William H. Winsborough. Automated trust negotiation
using cryptographic credentials. In ACM Conference on Computer and Communica-
tions Security, pages 46–57, 2005.

[29] Patrick Drew McDaniel and Atul Prakash. Methods and limitations of security policy
reconciliation. In IEEE Symposium on Security and Privacy, pages 73–87, 2002.

[30] Cisco Networks. http://netflow.cesnet.cz/n netflow.php.

[31] M. Pirretti, P. Traynor, P. McDaniel, and B. Waters. Secure Atrribute-Based Systems.

In ACM conference on Computer and Communications Security (ACM CCS), 2006.
To appear.

[32] V.V. Prasolov. Problems and Theorems in Linear Algebra. American Mathematical
Society, 1994.

[33] A. Sahai. Non-malleable non-interactive zero knowledge and adaptive chosen-

ciphertext security. In IEEE Symposium on Foundations of Computer Science, 1999.

[34] A. Sahai and B. Waters. Fuzzy Identity Based Encryption. In Advances in Cryptology
– Eurocrypt, volume 3494 of LNCS, pages 457–473. Springer, 2005.

[35] A. Shamir. How to share a secret. Commun. ACM, 22(11):612–613, 1979.

[36] A. Shamir. Identity Based Cryptosystems and Signature Schemes. In Advances in

Cryptology – CRYPTO, volume 196 of LNCS, pages 37–53. Springer, 1984.

[37] Nigel P. Smart. Access control using pairing based cryptography. In CT-RSA, pages
111–121, 2003.

[38] Ting Yu and Marianne Winslett. A unified scheme for resource protection in automated
trust negotiation. In IEEE Symposium on Security and Privacy, pages 110–122, 2003.

22
A Construction for Any LSSS-realizable Access Structure
Consider an access structure for which there exists a linear secret-sharing scheme that
realizes it. It is known that for every LSSS realizable access structure, there exists a
monotone span program (MSP) that computes the corresponding boolean function and
vice versa [4]. Thus, such an access structure can be represented by a monotone span
program.
In an attempt to accommodate more general and complex access structures, we present
attribute based encryption for the class of access structures which can be represented by
polynomial size monotone span programs. The access structure will be represented in the
form of a monotone span program M̂ (M, ρ). Each row of M will be labeled by an attribute
~ i.
from U and ρ(i) will denote the label of ith row M

A.1 Our Construction

The data will be encrypted under a set of attributes γ. The user should be able to decrypt
the data if and only if he holds the keys for an MSP M̂ (M, ρ) such that fM (γ) = 1; here
fM denotes the function that M̂ computes.
Let G1 be a bilinear group of prime order p, and let g be a generator of G1 . Additionally,
let e : G1 × G1 → G2 denote the bilinear map. A security parameter, κ, will determine the
size of the groups. Our construction follows:

Setup Define the universe of attributes U = {1, 2, . . . , n}. Now, for each attribute i ∈ U,
choose a number ti uniformly at random from Zp . Finally, choose y uniformly at random
in Zp . The published public parameters PK are:

T1 = g t1 , . . . , T|U| = g t|U | , Y = e(g, g)y

The master key MK is:

t1 , . . . , t|U| , y
Encryption (m, γ, PK) To encrypt a message m ∈ G2 under γ, choose a random value
s ∈ Zp and publish the ciphertext as:
0
E = (γ, E = mY s , {Ei = Tis }i∈γ )

Key Generation (M̂ , MK, PK) Input to the key generation algorithm is an access
structure in the form of an MSP M̂ (M, ρ). Let the dimensions of M be d × `. Choose
a random vector ~u from Z`p such that ~1 · ~u = y, that is, u = (u1 , u2 , . . . , u` ) such that
P` ~
i=1 ui = y. For each row vector Mi , give the following secret value to the user:

M~ ·~
i u
tρ(i)
Di = g

The set of above secret values is the decryption key D.

23
Decryption (E, D, PK) To decrypt the ciphertext E = (γ, E 0 , {Ei }i∈γ ) with the decryp-
tion key D, proceed as follows. Since fM (γ) = 1, the span program M̂ accepts γ. Thus,
there exist coefficients αi ∈ Zp such that,
X
~ i = ~1
αi · M
ρ(i)∈γ

Now,

Y Y ~ ·~
Mi u
E0/ e(Di , Eρ(i) )αi = E0/ e(g tρ(i) , g stρ (i) )αi
ρ(i)∈γ ρ(i)∈γ
Y ~
= mY s / e(g, g)sαi ·Mi ·~u
ρ(i)∈γ
P
~ i )·~
= mY s /e(g, g)s·( ρ(i)∈γ αi ·M u

~
= mY s /e(g, g)s·(1·~u)
= me(g, g)sy /e(g, g)sy
= m.

A.2 Proof of Security

We prove that the security of our scheme in the attribute-based Selective-Set model reduces
to the hardness of the Decisional BDH assumption.

Theorem 3 If an adversary can break our scheme in the Attribute-based Selective-Set

model, then a simulator can be constructed to play the Decisional BDH game with a non-
negligible advantage.

Proof: Suppose there exists a polynomial-time adversary A, that can attack our scheme
in the Selective-Set model with advantage ². We build a simulator B that can play the
Decisional BDH game with advantage ²/2. The simulation proceeds as follows:
We first let the challenger set the groups G1 and G2 with an efficient bilinear map, e
and generator g. The challenger flips a fair binary coin µ, outside of B’s view. If µ = 0,
the challenger sets (A, B, C, Z) = (g a , g b , g c , e(g, g)abc ); otherwise it sets (A, B, C, Z) =
(g a , g b , g c , e(g, g)z ) for random a, b, c, z. We assume the universe, U, is defined.

Init The simulator B runs A. A chooses the set of attributes γ it wishes to be challenged
upon.

Setup The simulator sets the parameter Y = e(A, B) = e(g, g)ab . For all i ∈ U, it sets
Ti as follows: if i ∈ γ, it chooses a random ri and sets Ti = g ri (thus, ti = ri ); otherwise it
chooses a random βi and sets Ti = g bβi = B βi (thus, ti = bβi ).
It then gives the public parameters to A. Notice that from A’s view all parameters are
chosen at random as in the construction.

24
Phase 1 A adaptively makes requests for several access structures represented in the
form of MSP such that γ is accepted by none of them.
Suppose A makes a request for the secret key for an MSP M̂ (M, ρ) such that fM (γ) = 0.
Let the dimensions of M be d × `. Let Mγ be the submatrix of M consisting of only
those rows whose label is in γ. To generate the secret key, B has to fix a random vector
u = (u1 , u2 , . . . , u` ) such that ~1 · ~u = ab. To do that, first define a vector v = (v1 , v2 , . . . , v` )
such that vi = bλi , for λi chosen uniformly at random from Zp . Now, consider the following
proposition [3, 32],
Proposition 1 A vector ~π is independent of a set of vectors represented by a matrix N if
and only if there exists a vector w ~ = ~0 while ~π · w
~ such that N w ~ 6= 0.
Since ~1 is independent of Mγ , there exists a vector w ~ = ~0 and ~1 · w
~ such that Mγ w ~ 6=
0(= h, say). Such a vector can be efficiently computed [3, 32]. Let w = (w1 , w2 , . . . , w` ).
Finally define the vector ~u as follows:
P
ab − b `k=1 λk
~u = ~v + ψ w
~ where ψ =
h
Notice that,
P`
k=1 λk
ab − b
ui = vi + ψwi = bλi + · wi
h
Let Mj = (xj1 , xj2 , . . . , xj` ). Give the secret key for the row Mj as follows. If ρ(j) ∈ γ,
then: P`
φ1 xji λj
Dj = B where φ1 = i=1
rρ(j)
Note that φ1 is completely known. This is a legitimate key because,
P`
~ j · ~u ~ j · (~v + ψ w) ~ j · ~v + ψ(M ~ j · w) ~ j · ~v + ψ · 0
M M ~ M ~ M i=1 xji λi
= = = =b· = bφ1
tρ(j) tρ(j) tρ(j) tρ(j) rρ(j)
If ρ(j) ∈
/ γ, then:
P` P` P`
φ2 φ3 i=1 xji i=1 xji (hλi − k=1 λk )
Dj = A g where φ2 = , φ3 =
hβρ(j) hβρ(j)
Note that φ2 , φ3 are completely known. This is a legitimate key because,
P` ³ P ´ ÃP ! ÃP !
a− `k=1 λk P`
M~ j · ~u b x
i=1 ji λ i + h
`
x `
x (hλ − λ )
i=1 ji i=1 ji i k=1 k
= =a + = aφ2 +φ3
tρ(j) bβρ(j) hβρ(j) hβρ(j)

Therefore, the simulator is able to construct a private key for the MSP M̂ . Furthermore,
the distribution of the private key for M̂ is identical to that of the original scheme.

Challenge The adversary A, will submit two challenge messages m0 and m1 to the
simulator. The simulator flips a fair binary coin, ν, and returns an encryption of mν . The
ciphertext is output as:
0
E = (γ, E = mν Z, {Ei = C ri }i∈γ )

25
If µ = 0 then Z = e(g, g)abc . If we let s = c, then we have Y s = (e(g, g)ab )c = e(g, g)abc ,
and Ei = (g ri )c = C ri . Therefore, the ciphertext is a valid random encryption of message
mν .
0
Otherwise, if µ = 1, then Z = e(g, g)z . We then have E = mν e(g, g)z . Since z is
0
random, E will be a random element of G2 from the adversary’s view and the message
contains no information about mν .

Phase 2 The simulator acts exactly as it did in Phase 1.

Guess A will submit a guess ν 0 of ν. If ν 0 = ν the simulator will output µ0 = 0 to

indicate that it was given a valid BDH-tuple otherwise it will output µ0 = 1 to indicate it
was given a random 4-tuple.
As shown in the construction the simulator’s generation of public parameters and private
keys is identical to that of the actual scheme.
In the case where µ = 1 the adversary gains no information about ν. Therefore, we
have Pr[ν 6= ν 0 |µ = 1] = 12 . Since the simulator guesses µ0 = 1 when ν 6= ν 0 , we have
Pr[µ0 = µ|µ = 1] = 21 .
If µ = 0 then the adversary sees an encryption of mν . The adversary’s advantage in
this situation is ² by definition. Therefore, we have Pr[ν = ν 0 |µ = 0] = 21 + ². Since the
simulator guesses µ0 = 0 when ν = ν 0 , we have Pr[µ0 = µ|µ = 0] = 12 + ².
The overall advantage of the simulator in the Decisional BDH game is 21 Pr[µ0 = µ|µ =
0] + 12 Pr[µ0 = µ|µ = 1] − 12 = 12 ( 12 + ²) + 12 21 − 12 = 12 ².

B Proof of Large Universe Construction

We prove that the security of large universe construction in the attribute-based Selective-Set
model reduces to the hardness of the Decisional BDH assumption.

Theorem 4 If an adversary can break our scheme in the Attribute-based Selective-Set

model, then a simulator can be constructed to play the Decisional BDH game with a non-
negligible advantage.

Proof: Suppose there exists a polynomial-time adversary A, that can attack our scheme
in the Selective-Set model with advantage ². We build a simulator B that can play the
Decisional BDH game with advantage ²/2. The simulation proceeds as follows:
We first let the challenger set the groups G1 and G2 with an efficient bilinear map, e
and generator g. The challenger flips a fair binary coin µ, outside of B’s view. If µ = 0,
the challenger sets (A, B, C, Z) = (g a , g b , g c , e(g, g)abc ); otherwise it sets (A, B, C, Z) =
(g a , g b , g c , e(g, g)z ) for random a, b, c, z.

Init The simulator B runs A. A chooses the challenge set, γ, an n element set of members
of Zp .

Setup The simulator assigns the public parameters g1 = A and g2 = B. It then chooses a
random n degree polynomial f (X) and calculates an n degree polynomial u(X) as follows:

26
set u(X) = −X n for all X ∈ γ and u(X) 6= −X n for some other X. Because, −X n and
u(X) are two n degree polynomials, they will have at most n points in common or they
are the same polynomial. This construction ensures that ∀X, u(X) = −X n if and only if
X ∈ γ.
u(i)
Now, the simulator sets ti = g2 g f (i) for all i = 1 to n + 1. Because f (X) is a random
n degree polynomial, all ti will be chosen independently at random as in the construction.
in +u(i) f (i)
Implicitly we have: T (i) = g2 g .

Phase 1 A adaptively makes requests for several access structures such that γ passes
through none of them. Suppose A makes a request for the secret key for an access structure
T where T (γ) = 0. To generate the secret key, B has to fix a polynomial Qx of degree dx
for every non-leaf node in the access tree.
Recall the functions PolySat and PolyUnsat from the proof of small universe case. The
simulator simply runs PolyUnsat(T , γ, A) on the main tree T . This defines a polynomial
qx for each node x of T such that qr (0) = a. Furthermore, for each leaf node x of T , we
know the polynomial qx completely if x is satisfied, and at least g qx (0) if x is not satisfied.
Simulator now defines the main polynomial Qx (·) = qx (·) for each node x of T . Notice
that this sets y = Qr (0) = a. The key corresponding to each leaf node x is given using its
polynomial as follows. Let i = att(x).

• If i ∈ γ.
Q (0) q (0)
Dx = g2 x T (i)rx = g2x T (i)rx
Rx = g rx
Where rx is chosen at random from Zp .

/ γ. Let g3 = g Qx (0) = g qx (0) .

• If i ∈
−f (i)
n +u(i) in +u(i) f (i) rx0
Dx = g3i (g2 g )
−1
in +u(i) 0
Rx = g3 g rx

Where rx0 is chosen at random from Zp .

The value in + u(i) will be non-zero for all i ∈

/ γ. This follows from our construction of
u(x). Above key components are legitimate because if we set rx = rx0 − inq+u(i)
x (0)
then,

−f (i)
n +u(i) in +u(i) f (i) rx0
Dx = g3i (g2 g )
−qx (0).f (i)
in +u(i) 0
= g in +u(i) (g2 g f (i) )rx
−qx (0)
q (0) in +u(i) f (i) in +u(i) f (i) rx0
= g2x (g2 g ) in +u(i) (g2 g )
q (0)
q (0) in +u(i) f (i) rx0 − in x
= g2x (g2 g ) +u(i)

q (0)
= g2x (T (i))rx
Q (0)
= g2 x T (i)rx

27
and,
−1 q (0)
n +u(i) 0 rx0 − in x
Rx = g3i g rx = g +u(i) = g rx
Therefore, the simulator is able to construct a private key for the access structure T .
Furthermore, the distribution of the private key for T is identical to that of the original
scheme.

Challenge The adversary A, will submit two challenge messages m0 and m1 to the
simulator. The simulator flips a fair binary coin ν, and returns an encryption of mν . The
ciphertext is output as:
0
E = (γ, E = mν Z, E 00 = C, {Ei = C f (i) }i∈γ )

If µ = 0 then Z = e(g, g)abc . Then the ciphertext is:

0
E = (γ, E = mν e(g, g)abc , E 00 = g c , {Ei = (g c )f (i) = T (i)c }i∈γ )

This is a valid ciphertext for the message mν under the set γ.

0
Otherwise, if µ = 1, then Z = e(g, g)z . We then have E = mν e(g, g)z . Since z is
0
random, E will be a random element of G2 from the adversaries view and the message
contains no information about mν .

Phase 2 The simulator acts exactly as it did in Phase 1.

0 0 0
Guess A will submit a guess ν of ν. If ν = ν the simulator will output µ = 0 to
0
indicate that it was given a valid BDH-tuple otherwise it will output µ = 1 to indicate it
was given a random 4-tuple.
As shown in the construction the simulator’s generation of public parameters and private
keys is identical to that of the actual scheme.
In the case where µ = 1 the adversary gains no information about ν. Therefore, we
0 0 0
have Pr[ν 6= ν |µ = 1] = 12 . Since the simulator guesses µ = 1 when ν 6= ν , we have
0
Pr[µ = µ|µ = 1] = 21 .
If µ = 0 then the adversary sees an encryption of mν . The adversary’s advantage in
0
this situation is ² by definition. Therefore, we have Pr[ν = ν |µ = 0] = 12 + ². Since the
0 0 0
simulator guesses µ = 0 when ν = ν , we have Pr[µ = µ|µ = 0] = 12 + ².
0
The overall advantage of the simulator in the Decisional BDH game is 21 Pr[µ = µ|µ =
0
0] + 12 Pr[µ = µ|µ = 1] − 21 = 12 ( 21 + ²) + 12 12 − 12 = 12 ².

28