Towards secure cloud data management

Towards Secure Cloud Data Management
Wenchao Zhou∗ William R. Marczak† Tao Tao∗ Zhuoyao Zhang∗ Micah Sherr∗
Boon Thau Loo∗ Insup Lee∗
∗ †
University of Pennsylvania University of California at Berkeley

ABSTRACT potential threats posed by malicious entities within the cloud infras-
This paper explores the security challenges posed by data-intensive tructure; (ii) we then discuss data management challenges in the ar-
applications deployed in cloud environments that span adminis- eas of secure distributed data processing, end-to-end query results
trative and network domains. We propose a data-centric view of verification, and trust policy management; (iii) we propose solu-
cloud security and discuss data management challenges in the ar- tions to many of the identified challenges using Declarative Secure
eas of secure distributed data processing, end-to-end query result Distributed Systems (DS2 ) [26], a declarative platform for speci-
verification, and cross-user trust policy management. In addition, fying, analyzing, and deploying secure information systems; and
we describe our current and future efforts to investigate security finally, (iv) we outline our current and future efforts towards re-
challenges in cloud data management using the Declarative Secure solving the remaining cloud data management security challenges.
Distributed Systems (DS2 ) platform, a declarative infrastructure for
specifying, analyzing, and deploying secure information systems.
2. THREATS AND CHALLENGES
1. INTRODUCTION We categorize the threats and challenges of secure cloud data
Cloud computing is a popular computing paradigm in which com- management in terms of the three components of most cloud com-
putation is moved away from a personal computer or an individ- puting systems: infrastructure, communication network, and users.
ual application server to a “cloud” of computers in the network.
Cloud computing holds the promise of revolutionizing the manner Infrastructure: The cloud infrastructure – i.e., the hosts that pro-
in which enterprises manage, distribute, and share content. By out- vide cloud computing and data-store services – can be compro-
sourcing web hosting and other IT capabilities to one or more cloud mised by a malicious adversary. An attacker who controls nodes
providers, enterprises may achieve significant cost savings. on the cloud can launch a variety of attacks, including generating
In this paper, we propose a data-centric view of cloud secu- incorrect query results, conducting denial-of-service (DoS) attacks,
rity – one that focuses on the needs and challenges posed by data- and gaining unauthorized access to sensitive information.
intensive applications deployed in cloud environments than span In addition to securing the nodes that constitute the cloud, the
administrative and network domains. Cloud security issues have re- distributed data processing that makes the cloud useful must also
cently gained traction in the research community, where the focus be secured. For instance, in an untrusted setting, an operation that
has primarily been on securing the low-level operating systems or spans across multiple nodes should be authenticated (and possibly
virtual machine implementations (e.g., [22]) on which cloud com- encrypted). All forms of access control and authorizations of ma-
puting services are deployed. chines and users should be seamlessly integrated with the data pro-
Our work complements such existing research and further intro- cessing layer of the cloud. Ideally, users of the cloud should be able
duces a data-driven approach towards cloud security. Our work to validate the completeness and integrity of any query processing
is motivated by the observation that the majority of applications that the cloud has performed.
deployed on cloud infrastructures will involve some form of dis- Communication Network: Cloud infrastructures can span across
tributed management, typically in the form of database backends multiple network domains (e.g., across geographically distributed
that are partitioned and replicated. A data-centric view is therefore data centers), or in the case of cloud federations [4], the cloud itself
well-suited for analyzing the security properties of such data-driven may consist of multiple providers that cross administrative bound-
cloud computing applications. Further, our data-driven approach aries. Communication between cloud nodes must be properly se-
enables us to adopt from the ground up well-explored techniques cured to protect the confidentiality and integrity of data.
from the database community, such as recent work on query results Users: Security mechanisms must ensure that users of the cloud
verification [19, 20, 9, 17] of outsourced databases. may access only data that they are authorized. In our context, users
Moreover, as the Internet continues to evolve, it is increasingly refer to any customer of the cloud provider (for example, an enter-
dominated by applications that require integration, sharing, and in- prise that outsources its database-intensive operations to the cloud).
teroperable communication among various services and users on We argue that the existing model of isolating users within virtual
the Internet. Such applications include traditional retail portals that machines or slices is too restrictive. In several cloud applications
integrate content for comparison shopping and social-network sys- (e.g., mashups, collaborative analysis, social networks), data shar-
tems that incorporate mashups, blogging, social messaging, and ing is an integral aspect, and any cloud data management middle-
more recently, collaborative social analysis. A comprehensive cloud ware should provide a framework for securely sharing data between
security solution has to provide mechanisms for securely sharing users. Since users can dynamically enter and leave the cloud, access
data among users and content providers, even across administrative control policies should be extensible and easily modifiable.
and enterprise boundaries. A data-centric approach enables us to
build upon well-studied database techniques, ranging from database Given the above security challenges, we next describe DS2 in
access control [21] to recent work at using query engines to enforce terms of its declarative language and capabilities. In Sections 4
and implement extensible trust management policies [16, 26]. through 6, we describe how DS2 makes significant strides towards
This paper makes the following contributions: (i) We categorize secure cloud data management.

1

3. DS2 PLATFORM physical address (e.g., IP addresses) or a logical address (e.g., over-
2 lay identifier). In a multi-user multi-layered network environment
DS is a declarative platform for specifying, implementing, and
deploying networked information systems. In prior work [26], we where multiple users and overlay networks may reside on the same
demonstrate the flexibility and compactness of the Secure Network physical node, S can include a username and network identifier.
Datalog (SeNDlog) language via secure specifications of the path- Import/export predicates: The SeNDlog language allows differ-
vector routing protocol, Chord distributed hash table (DHT) [24], ent principals or contexts to communicate by importing and export-
and the PIER [10] distributed query processor. DS2 enhances SeND- ing tuples. The communication serves two purposes: (1) to dissem-
log and its distributed execution engine by providing additional inate update and maintenance messages and (2) to distribute the
support for dynamically layering multiple overlays at runtime (e.g. derivation of security decisions.
PIER over Chord) [15] and enabling security policies to be enforced During the evaluation of SeNDlog rules, derived tuples can be
and integrated across various layers. communicated among contexts via the use of import predicates and
In DS2 , network protocol and security policies are specified in export predicates. An import predicate is of the form “N says
SeNDlog [26], a declarative language primarily rooted in DataLog p” in a rule body, indicating that principal N asserts the predicate p.
that unifies declarative networking [14, 13] and logic-based access The use of export predicates ensures confidentiality and prevents
control [2] specifications. SeNDlog programs are disseminated and information leakage by exporting tuples only to specified princi-
compiled at each node into individual execution plans. When exe- pals. An export predicate is of the form “N says p@X” in a rule
cuted, these execution plans both implement the specified network head, where principal N exports the predicate p to the context of
protocol as well as enforce its desired security policies. As a core principal X. Here, X can be a constant or a variable. If X is a vari-
feature of declarative networking, SeNDlog requires orders of mag- able, the variable X is required to occur in the rule body. In rule
nitude less code than imperative languages [14]. r2 of the above example, node S will export reachable tuples to
SeNDlog further extends the basic declarative networking lan- node D. As a shorthand, “N says” can be omitted if N is where the
guage by adding support for authenticated communication. SeND- rule resides.
log integrates two commonly used constructs in distributed trust The says operator implements an authentication scheme [12]
management languages (e.g., Binder [8]): (1) the notion of con- that allows the receiver of the tuple to verify its authenticity. The
text to represent a principal in a distributed environment and (2) a implementation of says may depend on the system and its con-
distinguished operator says that abstracts away the details of au- text. In an untrusted environment, says may require digital signa-
thentication [8, 12]. tures. In a trusted environment, says may simply append a cleart-
To demonstrate the key language features of SeNDlog, we present ext principal header to a message (a less computationally expensive
the specification of a simple example of a distributed program that operation than using digital signatures). Somewhere in between,
computes all pairs of reachable nodes: cryptographic signatures may be applied only to certain important
messages or when communicating with specific principals. DS2
At S:
r1 reachable(S,D) :- link(S,D). provides sufficient flexibility to support a variety of cryptographic
r2 reachable(D,Z)@D :- link(S,D), W says reachable(S,Z). authentication primitives.
In addition to authentication, DS2 also provides mechanisms to
The program consists of two rules (r1 and r2) that are executed ensure the confidentiality of transmitted tuples. DS2 may optionally
in the context of node S. The program computes the set of reachable encrypt transmitted messages to ensure that they cannot be read by
nodes (reachable) given the set of network links (links) between adversaries or other unauthorized parties.
S and D. Rule r1 takes link(S,D) tuples, and computes single-hop
reachability reachable(S,D). In rule r2, multi-hop reachability 4. SECURE QUERY PROCESSING
is computed. Informally, rule r2 means that “if there is a link be- A significant challenge in cloud data management is ensuring
tween S and D1 , and there exists a node W asserting that Z is reach- that all query processing is carried out securely within a cloud in-
able from S, then D can also reach Z by using S as the intermediate frastructure. To be secure, query processing must (1) authenticate
node.” The says primitive in rule r2 specifies that the authentic- users and machines involved in query processing, (2) secure the
ity of the received reachable tuple should be checked, ensuring it transfer of data across machines in the cloud, and (3) ensure the in-
indeed comes from W as it claims. tegrity of all query results. All three requirements can be directly
By modifying the above SeNDlog rules, one can easily develop applied to mitigate potential threats (see Section 2) at the infrastruc-
more complex routing protocols and policies (e.g., secure BGP). ture, network, and user levels.
Since DS2 ’s dataflow framework captures information flow as dis- Along an orthogonal axis, there is also the issue of where the
tributed queries, data provenance [3] may be utilized to securely security guarantees are enforced. The obvious solution is to en-
“explain” the existence of any network state (analogous to the use force security within the cloud infrastructure. This places security
of proof-trees [11, 6] in security audits). burdens entirely on the cloud provider, consequently requiring all
We describe two language features of SeNDlog of relevance to clients to trust their providers. Alternatively, end-to-end verifica-
security, with additional details in [26]: tion approaches allow users to verify results without requiring the
Communication context: Due to the distributed nature of network implicit trust of any given cloud provider, but do so at the cost of
queries, a principal does not have control over rule execution at more complex application design.
other nodes. SeNDlog achieves secure distributed query processing In this section, we focus on the former in-cloud mechanisms and
by allowing programs to interoperate correctly and securely via the defer our discussion of the latter approach to Section 6. We con-
export and import of rules and derived tuples across contexts. centrate our discussion on an example based on an authenticated
In the above example, the rules are in the context of S, where S implementation of MapReduce [7].
is a variable that is assigned upon rule installation. In a distributed
environment, S represents the network address of a node: either a 4.1 Example: Authenticated MapReduce
MapReduce [7] is a paradigm that has been extensively used in
1
Bidirectional links are assumed, where D has a link to S as well. cloud applications for efficiently utilizing parallel resources to solve

2

certain classes of distributed problems. Generally, a MapReduce (MW) share a common key to sign tuples. Reduce workers verify
job consists of two steps: map and reduce. At the map step, a tuples as having been produced by the node MW.
master node splits the job into small sub-problems and distributes Alternatively, authentication can occur at the user level. User-
them to map workers. The results produced by the map workers are level authentication is useful when intermediate and/or final com-
then collected at the reduce step, and are combined into a solution putation results are shared among users. For instance, a user alice
to the original problem. that runs applications on the cloud may only trust input data from a
In the MapReduce paradigm, users supply the Map and Reduce particular user bob. Here, the owner of a MapReduce job (e.g. bob
functions; the middleware transparently handles distribution and in the above example) signs each tuple with its own signature. If
fault tolerance. In an untrusted environment, all participating nodes the number of workers is large, key management is nontrivial. DS2
in a MapReduce job must be authenticated. Once authenticated, supports public key infrastructures (e.g., certificate authorities) to
additional access policies can be enforce (see Section 5). enable workers to more easily generate and verify signatures.
As an example, we will consider the WordCounting program in
which MapReduce is used to count the occurrences of words in 4.2 Evaluation
webpages. The following rules (m1-m2 for map steps and r1-r2 This section presents an evaluation of DS2 . The goal of our
for reduce steps) demonstrate an authenticated implementation of evaluation is two-fold: (1) to experimentally validate DS2 ’s abil-
MapReduce written in SeNDlog and executed using DS2 : ity to implement secure cloud applications using the MapReduce
At MW: paradigm, and (2) to study the additional overhead incurred by adopt-
m1 map(ID,Content) :- file(MW,ID,Content). ing authentication in the system.
m2 emits(MW,Word,Num,Offset)@RW :- word(Word,Num,Offset),
reduceWorker(RID,RW), RID==f_SHA1(Word). As a workload, we adapt the WordCounting example from the
previous section, implemented as a two-stage MapReduce job:
At RW:
r1 reduceTuple(Word,a_LIST<Num>) :-
NormalizeStage: In the first stage, the master node distributes a set
MW says emits(MW,Word,Num,Offset). of webpages to the map workers. Map workers filter out all HTML
r2 reduce(Word,List) :- reduceTuple(Word,List), tags. The cleansed webpages are further split into small chunks
Master says rBegin(RW).
and submitted to the reduce workers. Each reducer distributes the
In the program shown above, rules m1 and m2 are within the con- webpage chunks to the map workers of the next stage (CountStage)
text of a map worker MW, and rules r1 and r2 are in the context of based on SHA-1 digests of the chunks.
a reduce worker RW. As input to the MapReduce operation, the user CountStage: The MapReduce operation at the second stage per-
supplies the Map and Reduce functions which encode the detailed forms the WordCounting operation. Map workers take as input the
operations in the form of imperative programs. We provide a brief small chunks produced from the previous stage and emit ( WORD ,1)
description of these functions: pairs to the reduce workers. The reduce workers aggregate such
Map operation: Rule m1 takes as input the file predicates which pairs and generate the result: the occurrence counts of each word in
contain the identifier and content of each document. The rule gen- the input webpages.
erates the (ID, Content) pairs – i.e., the local map tuples – which
As the input of the MapReduce job, we randomly selected 6,400
are then passed to instances of the user-defined Map function. Upon
webpages from a crawled dataset belonging to the Stanford Web-
receiving an (ID, Content) pair, each Map instance operates as fol-
Base project [1]. The average size of a webpage is 6KB.
lows. It splits the contents of a document (stored in the Content
We perform the experiments within a local cluster of 16 quad-
field) into separate words. For each word, a ( WORD ,1) pair is gen-
core machines with Intel Xeon 2.33GHz CPUs and 4GB RAM run-
erated (stored in word tuples), denoting that the occurrence count
ning Linux 2.6, interconnected by Gigabit Ethernet. We deploy 16
of the word should be increased by one. To prevent miscounting the
map workers and 16 reduce workers on the NormalizeStage and
occurrence of a word that appears at multiple places within the same
32 map workers and 128 reduce workers on the CountStage. Each
document, the Offset of each word in the document is tagged with
physical machine runs a total of 12 MapReduce worker instances.
the corresponding word tuple, which is then sent back to MW as the
To avoid packet-drops due to congestion, we rate-limit the number
result of the map function.
of packets sent per second.
Rule m2 takes as input word tuples and distributes them (in the
To evaluate the overhead incurred by performing authentications,
form of emits tuples) to reduce workers. word tuples are assigned
we constructed three versions of the WordCounting job: NoAuth,
to reduce workers based on SHA-1 hashes of the words. Addition-
RSA-1024 and HMAC. In NoAuth, MapReduce workers transmit tu-
ally, to leverage access control in a multi-user setting, a signature is
ples without including the sender’s signature, whereas in RSA-1024
included within each emits tuple using the says primitive.
and HMAC the communication between different map and reduce
Reduce operation: Reduce workers receive and authenticate (e.g., workers are authenticated using 1024-bit RSA signatures and 160-
via digital signatures) emits tuples from map workers. The tuples bit SHA-1 HMACs, respectively.
are then grouped by the key field word, as shown in rule r1. The Figure 1 shows the per-node bandwidth usage over time achieved
a LIST aggregate operator maintains the occurrences of each word by the three versions of the WordCounting job. All three versions
in a list structure. incur spikes in their bandwidth utilization in the first 30s. The
After the map workers complete their job, a master node sends a spikes are mainly attributed to the cross-node communication at the
rBegin tuple to each reduce worker, signaling that they should start NormalizeStage. The MapReduce operation at this stage is com-
the computation of the reduce job. reduce tuples are sent to the putationally cheap but network intensive, as the tuples transmitted
user-defined Reduce instances, each of which takes as arguments in this phase consist of relatively large chunks of documents (as
a word and the list of its occurrences. Based on these lists, the compared to ( WORD ,1) pairs in the CountStage).
Reduce instances generate and emit the final results – i.e., the total We observe that NoAuth finishes the computation in 350 seconds,
occurrence counts of words. whereas HMAC and RSA-1024 incur an additional 17.4% (60s) and
Node- vs. user-level authentication: The above program enforces 78.3% (270s) overhead in query completion latency, respectively.
authentication at the node level: map workers on the same node The increase in query completion time is due primarily to the com-

3

putation incurred by signature generation and verification. query rewriting and other techniques. Our approach builds upon
Given that a dominant portion of the communication is small- traditional database-based and logic-based access control mecha-
sized ( WORD ,1) pairs, the relative overhead incurred by tagging nisms for reasoning about trust policies.
signatures with transmitted tuples is significant. We observe that
the respective aggregated communication overheads of HMAC and 5.1 View-based Access Control and Rewrites
RSA-1024 are 18.4% (7.5MB) and 53.3% (21.8MB) higher than Traditional databases typically utilize view-based access control
NoAuth. However, due to the use of network throttles in our evalu- in which views are expressed in SQL, with access controls to these
ation, the per-node bandwidth utilization for the three versions are views enforced via explicit permissions granted to authorized users.
similar. Such views are easily expressible in DS2 . For example, if a user
alice owns an employees predicate that stores the names, de-
4.3 Layered Encryption and Authentication partments, and salaries of employees, then she may create a security
The above example illustrates authenticated computations at the view for some other user, Bob:
data processing layer of a cloud infrastructure. Given that cloud in- At alice:
frastructures can span multiple network or administrative domains, sv1: employee_sv_bob(Name,Dept) :-
communication must also be properly secured. SeNDlog provides employee(Name,Dept,Salary), Salary < 5000.
sv2: predsecview("employee","employee_sv_bob",bob).
additional security constructs for encrypting communication.
Authentication and encryption can be enforced either at the data In this example, alice allows bob to view only those employ-
processing layer or at the network layer. Since DS2 is based on ees in any department who have a salary less than $5000. The
declarative networking, it is a natural platform for implementing employee sv bob predicate represents both a horizontal
secure communication substrates (e.g., implementing secure rout- (salary<5000) and vertical (salary column omitted) partition of the
ing protocols on the control plane and secure forwarding on the employee predicate. Note that employee sv bob (defined by rule
data plane). For instance, when the map and reduce workers trans- sv1) is dynamically applied as the predicates in the rule body up-
mit tuples, these tuples may traverse untrustworthy channels. DS2 date, and hence its contents change with the contents of the
provides a mechanism for performing secure forwarding of tuples, employee predicate. In rule sv2, an additional predicate
effectively constructing a transient virtual private network between predsecview is maintained to associate security views with the
workers. For example, the following SeNDlog program performs protected predicate.
authenticated packet forwarding given the forwarding table at ev- DS2 can easily express security views that apply to entities other
ery node in the network: than single users. For example, alice may delegate access to the
At Router, users whom a certificate authority believes are good by replacing
f1 packet(Pid,Dest,Data)@NextHop :- rule sv2 with the following rule:
forwarding(Dest,NextHop), Router != Dest,
P says packet(Pid,Dest,Data). sv2’: predsecview("employee","employee_sv_good",U) :-
cert_authority says good(U).
The forwarding table maintains the nextHop router along the
shortest path to any given destination (Dest). Rule f1 forwards In addition to supporting simple horizontal and vertical slicing,
a packet Pid received from the upstream router P along the best DS2 also permits more complex partitioning of access control. For
path to Dest via the nextHop router. This payload is recursively example, access rights may be based on provenance information
routed by rule f1 to the destination, with authentication occurring at (e.g., Alice allows Bob to view only those employee tuples derived
every hop. Similar declarative techniques can be used to compactly using information from Bob). In general, SeNDlog’s flexibility en-
specify encrypted communication using the SeNDlog language. ables DS2 to support a wide variety of access control policies.
Most importantly, DS2 ’s ability to unify different layers of pro- Enforcement. We have thus far left unspecified how security views
tocols (i.e., data processing and network routing) make it an inter- are enforced. Standard Datalog cannot prevent users from submit-
esting framework for performing cross-layer protocol analysis. For ting a query that references predicates directly, nor can unmodified
example, DS2 may be well-suited for constructing provenance in- Datalog dynamically rewrite queries to refer to the security views.
formation [18] that spans system and network layers and crosses To enable these capabilities, we extend Datalog to represent the cur-
administrative domains. rently executing queries in tables accessible to the program – a con-
cept we call the DS2 meta-model. Our approach is similar to the
5. SECURE DATA SHARING recently proposed Evita-Raced meta-compiler catalog [5].
Data sharing and integration are integral to many classes of appli- Preventing queries from directly reading predicates (e.g.,
cations that operate in a multi-user cloud environment. If the cloud employee) is made possible by adding schema constraints called
spans administrative domains, a domain may wish to define poli- meta-constraints [16] to the meta-model. Meta-constraints restrict
cies to share data with another domain for the purposes of further the set of allowable queries. In particular, DS2 uses meta-constraints
distributing or outsourcing of computation. Moreover, applications to express that users can only insert queries that refer to security
deployed in the cloud may be isolated within their own virtual ma- views:
chines. Such isolation may be relaxed by enabling principals to
says(U,R), body(R,A), functor(A,P) -> predsecview(_,P,U).
securely interact using access control policies.
The enforcement of access control policies can become compli- The above example introduces the schema constraint format. Note
cated by the fact that users may frequently enter or leave the cloud, that instead of :-, the head and body are separated by a right-arrow
requiring policies to be rapidly updated. Real-world trust policies (->). The logical meaning of this format is that if, for any assign-
are complex and may require functionality beyond that offered by ment of the variables, the left hand side (LHS) is true, then the right
traditional tuple-level access control policies. Devising a scheme hand side (RHS) must also be true. If, on the other hand, the LHS
to respond to the dynamics and complexity of cloud environments is true but the RHS is false, then evaluation of the query terminates
represents a significant challenge. with an error.
In this section, we propose methods in which DS2 can be em- In the above constraint, says associates a user U with a query R,
ployed to readily support complex and dynamic policies through and body and functor are meta-model predicates that examine the

4

700
NoAuth h1~8 h1~8
Per‐Node Bandwidth (KBps)
e Bandwidth (KBps)
600
RSA‐1024
500
HMAC h1~4 h5~8 h1~4 h5~8
400

300

200 h1~2 h3~4 h5~6 h7~8 h1~2 h3~4

100

0
0 80 160 240 320 400 480 560 640 h1 h2 h3 h4 h5 h6 h7 h8 h1 h2 h3 h4
Time Elapsed (s)

Figure 1: Bandwidth utilization (KBps) Figure 2: MHT for table X Figure 3: P-MHT for table X1

structure of the rule. The constraint expresses that every predicate their database operations to third parties (e.g., cloud operators).
mentioned in the body must be the security view of some other End-to-end verification works by requiring nodes to transmit ver-
predicate (since the variable for this other predicate is unimportant ification objects (VOs) along with query results. Clients use VOs to
in the constraint, we represent it with an underscore). verify the correctness of the tuples in the returned resultsets.
Rewriting queries is equally simple using meta-rules: Below, we briefly review existing end-to-end verification tech-
-rulebody(R,B), +rulebody(R,C), +atom(C), +atompred(C,S), niques and discuss some of their limitations. Additionally, we de-
+atomargs(C,N,V) :- scribe our proposed adaptations of these techniques for the cloud
P says R, rulebody(R,B), atompred(B,P), environment. For ease of exposition, we consider the case in which
predsecurityview(P,S), predargs(S,N,V).
a content provider outsources its database operations to a third-
This rule leverages DS2 ’s support for code generation (also called party server. The provider (the client) wishes the verify the cor-
updating rules). Code generation changes the current state of the rectness of the query results.
database by asserting and retracting specific tuples. In this case,
the rule updates the body of the query by retracting (-) the body 6.1 Existing Techniques
atom representing the original predicate and asserting (+) the body A simple approach for creating a VO is via per-tuple
atom representing the security view. Note that the logical meaning signatures [19]. Here, the data owner digitally signs each tuple be-
of a conjunction of atoms in the head of a rule q1,q2,...,qm :- fore storing them on the server’s site. In response to a query result,
p1,p2,...,pn is the set of m rules: qi :- p1,p2,...,pn. the server sends the matching tuples and their corresponding signa-
tures to prove the integrity and correctness (but not necessarily the
5.2 Multi-user Multi-stage MapReduce completeness) of the results.
In Section 4, we introduced an authenticated version of MapRe- A more robust approach that ensures completeness of query re-
duce. Given that both access control policies and the MapReduce sults relies on signature-chaining [20] in which a signature is gen-
program are specified within SeNDlog, it is easy to integrate access erated from each tuple and its two adjacent tuples (neighbors) in
control policies as additional rules in the program. The program- the table. The signature chain guarantees completeness since it is
mer needs only to supply a set of policy rules; DS2 automatically not possible to omit a result tuple while still satisfying the neigh-
performs the query rewrites to generate the appropriate rules for bors’ signatures. The order of tuples in the table can be determined
executing the MapReduce program. based on tuple ID or a sort on one of the attributes. Upon com-
In practice, given that MapReduce operations can span multiple puting the query results, the server returns the result tuples, their
stages (i.e., compose several MapReduce operations), access con- neighbors, and the corresponding signatures. The approach may
trol can be enforced across stages. For instance, if alice starts a be customized for range queries to ensure that a malicious server
MapReduce operation and the resulting output is used by bob for cannot omit a result tuple within the queried range.
his subsequent MapReduce operation, one needs to ensure that bob Unfortunately, the signature-chaining scheme is inefficient in
only accesses data that he is authorized to view. One na¨ve solution
ı terms of storage, communication, and computation costs. An al-
is to enforce access control only at the boundaries of each MapRe- ternative authentication approach for range queries uses a Merkle
duce operation – i.e., on the resulting output between jobs. A more Hash Tree (MHT) [9, 17]. The MHT is constructed as a binary tree
sophisticated solution that we are currently exploring is to “push- in a bottom-up fashion, in which the leaves correspond to the hashes
down” selection predicates dynamically into each MapReduce ex- of the tuples in a sorted table. Non-leaf nodes correspond to the
ecution plan. This ensures that proper filtering is applied early so hash of the concatenation of the children, and the tree construction
that only authorized data flows downstream to later stages of the process is carried out recursively to the root of the tree. The con-
MapReduce operations. In this example, the trust relationships be- tent provider maintains the MHT along with the master database.
tween alice and bob may result in filtering of input data as early When it outsources the database to servers in the cloud, it sends
as alice’s MapReduce operation, causing only the data that bob both the database and the MHT to the server. In addition, it also
is authorized to view to be processed. We are currently investigat- authenticates the MHT root with its signature. A VO consists of
ing methods to adapt the dynamic rewrite capabilities presented in the signed root and the collection of log(n) internal tree nodes that
[16] to handle more complex sharing, particularly across users and enable the verifier to recompute the root of the MHT. In the verifica-
multiple stages of MapReduce computations. tion process, given a set of query results and the corresponding VO
returned by the server, a client verifies the query results by comput-
6. END-TO-END VERIFICATION ing the MHT based on the query results. If the root of its computed
The above discussion assumes that security is enforced within MHT matches that of the signed root contained in the VO, then the
the cloud infrastructure by the cloud provider. That is, the provider query results are guaranteed to be valid and complete.
correctly applies the security mechanisms described in Section 4 to Built upon the techniques presented above for verifying range
support authenticated and confidential queries. queries, recent work [25] proposes three techniques – AISM, AIM,
Alternatively, queries may be protected using end-to-end verifi- and ASM – to perform join operations with the capability to verify
cation. These approaches are motivated by clients who outsource the completeness and integrity of the computation results. While

5

AIM utilizes the MHTs of the both joining tables, AISM and ASM provides query authenticity with little overhead. We plan to deploy
relax the requirement on the availability of MHTs by sacrificing our prototype on PlanetLab to validate its performance in more re-
performance. alistic settings.
The SeNDlog language and interpreter are available for download
6.2 End-to-End Verification in the Cloud at http://netdb.cis.upenn.edu/rapidnet/downloads.html.
Several characteristics of the previously described existing ap- Although we have presented our examples as compact SeNDlog
proaches make them unsuitable for the cloud environment: First, programs, other policy specification languages may also be appli-
the requirement of a master database at the content provider limits cable. As an alternative that achieves backward compatibility and
the possibility of queries that integrate content from multiple clients legacy support, we are exploring utilizing DS2 as a separate exten-
on the cloud. Second, the MHT-based approach is not amenable to sible trust management policy engine that may be incorporated into
a database in which the contents are dynamic, since the MHT needs existing distributed computing software (e.g. Hadoop).
to be recomputed whenever the base data changes. Third, in the
cloud environment, data can potentially be distributed across sev- 8. Stanford WebBase. http://diglib.stanford.edu:8091/ testbed/doc2/WebBase/.
REFERENCES
eral nodes, complicating the process of MHT construction.2 [1]
[2] A BADI , M. Logic in Access Control. In Symposium on Logic in Computer
We are currently exploring techniques that address all of the above Science (June 2003).
limitations by developing new capabilities using the DS2 platform. [3] B UNEMAN , P., K HANNA , S., AND TAN , W. C. Why and where: A
One of our ongoing research approaches is to add completeness and characterization of data provenance. In ICDT (2001).
integrity verification capabilities to SeNDlog’s query results. Addi- [4] C AMPBELL , R., G UPTA , I., H EATH , M., KO , S. Y., KOZUCH , M., K UNZE ,
M., K WAN , T., L AI , K., L EE , H. Y., LYONS , M., M ILOJICIC , D.,
tionally, we are exploring the application of data provenance for O’H ALLARON , D., AND S OH , Y. C. Open cirrus cloud computing testbed:
checking the completeness and integrity of data processing. Along Federated data centers for open source systems and services research. In
another line, researchers have been exploring challenge-based ap- HotCloud (2009).
proaches (see, for example, [23]) to verify the compliance of pro- [5] C ONDIE , T., C HU , D., H ELLERSTEIN , J. M., AND M ANIATIS , P. Evita raced:
Metacompilation for declarative networks. In VLDB (2008).
gram executions to their expected behaviors. In general, we plan to [6] C ROSBY, S. A., AND WALLACH , D. S. Efficient Data Structures for
support distributed verification with the ability to handle dynamism Tamper-Evident Logging. In 18th USENIX Security Symposium (August 2009).
in both data and nodes within the DS2 system. [7] D EAN , J., AND G HEMAWAT, S. Mapreduce: Simplified data processing on
Due to space constraints, we focus below on the challenge of large clusters. In Proceedings of Usenix Symposium on Operating Systems
Design and Implementation (December 2004).
constructing MHT for distributed verification. Consider the case in [8] D E T REVILLE , J. Binder: A logic-based security language. In IEEE Symposium
which a table is horizontally partitioned across a set of machines, on Security and Privacy (2002).
where each machine maintains a range of tuples in a sorted table. [9] F EIFEI L I , M. H., AND KOLLIOS , G. Dynamic authenticated index structures
We propose an approach called partitioned MHT (P-MHT) that may for outsourced databases. In SIGMOD (2006).
[10] H UEBSCH , R., H ELLERSTEIN , J. M., L ANHAM , N., L OO , B. T., S HENKER ,
be applied to data partitions, thus allowing efficient data mainte- S., AND S TOICA , I. Querying the Internet with PIER. In VLDB (2003).
nance and verification across cloud providers. [11] J IM , T. SD3: A Trust Management System With Certified Evaluation. In IEEE
To demonstrate how P-MHT can be applied, we consider the ta- Symposium on Security and Privacy (May 2001).
ble X = {x1 , x2 , x3 , ..., x8 }. The MHT of the table is shown [12] L AMPSON , B., A BADI , M., B URROWS , M., AND W OBBER , E. Authentication
in Distributed Systems: Theory and Practice. ACM TOCS (1992).
in Figure 2, where h(xi ) is the digest of the tuple xi . Suppose
[13] L OO , B. T., C ONDIE , T., H ELLERSTEIN , J. M., M ANIATIS , P., ROSCOE , T.,
X is partitioned into three sub-tables, i.e., X1 = {x1 , x2 , x3 }, AND S TOICA , I. Implementing Declarative Overlays. In ACM SOSP (2005).
X2 = {x4 , x5 , x6 }, and X3 = {x7 , x8 }. In addition to storing [14] L OO , B. T., H ELLERSTEIN , J. M., S TOICA , I., AND R AMAKRISHNAN , R.
the tuples, each sub-table maintains a P-MHT that maintains suffi- Declarative Routing: Extensible Routing with Declarative Queries. In
SIGCOMM (2005).
cient information for the local generation of the VO for any tuple
[15] M AO , Y., L OO , B. T., I VES , Z., AND S MITH , J. M. MOSAIC: Unified
in the sub-table. For instance, Figure 3 shows the P-MHT for X1 . Declarative Platform for Dynamic Overlay Composition. In ACM CoNEXT
Although the P-MHT is a partial MHT, it may still generate the VO 2008.
for any tuple maintained in X1 . For example, the VO for x3 is [16] M ARCZAK , W. R., Z OOK , D., Z HOU , W., A REF, M., AND L OO , B. T.
Declarative reconfigurable trust management. In 4th Biennial Conference on
h(x1−2 ), h(x4 ), h(x5−8 ), which, when combined with x3 , can be Innovative Data Systems Research (CIDR) (2009).
used to compute the MHT root and thus validate the integrity of the [17] M OURATIDIS , K., S ACHARIDIS , D., AND PANG , H. Partially materialized
query result by comparing with the root signed by the table owner. digest scheme: an efficient verification method for outsourced databases. VLDB
Our approach works for tables that are ordered and range par- Journal) (2009).
[18] M UNISWAMY-R EDDY, K., B RAUN , U., H OLLAND , D. A., M ACKO , P.,
titioned across nodes. Verifying query results across hash-based M ACLEAN , D., M ARGO , D., S ELTZER , M., AND S MOGOR , R. Layering in
partitions is still an open area of future work. provenance systems. In USENIX Annual Technical Conference (2009).
[19] M YKLETUN , E., NARASIMHA , M., AND T SUDIK , G. Authentication and
integrity in outsourced databases. In Symposium on Network and Distributed
7. CONCLUSION Systems Security (NDSS) (2004).
This paper outlines the challenges of secure cloud data manage- [20] PANG , H., AND TAN , K.-L. Verifying Completeness of Relational Query
ment. We propose security mechanisms based on the DS2 system Answers from Online Servers. ACM Transactions on Information and Systems
Security) (2008).
that address these challenges. Our initial results indicate that DS2 [21] R IZVI , S., M ENDELZON , A. O., S UDARSHAN , S., AND ROY, P. Extending
2 query rewriting techniques for fine-grained access control. In SIGMOD (2004).
The authenticated join processing techniques presented in the pre-
[22] S ANTOS , N., G UMMADI , K. P., AND RODRIGUES , R. Towards Trusted Cloud
vious section are also also inapplicable when tables are partitioned Computing. In HotCloud (2000).
across nodes: AISM and ASM require on-the-fly sort on at least one [23] S ION , R. Query execution assurance for outsourced databases. In VLDB (2005).
joining table. Such a requirement lacks an efficient distributed solu- [24] S TOICA , I., M ORRIS , R., K ARGER , D., K AASHOEK , M. F., AND
tion and seriously jeopardizes the performance and even feasibility BALAKRISHNAN , H. Chord: A Scalable Peer-to-Peer Lookup Service for
of these two approaches. On the other hand, the AIM algorithm Internet Applications. In SIGCOMM (2001).
is still applicable even when tables are distributed across nodes; [25] YANG , Y., PAPADIAS , D., PAPADOPOULOS , S., AND K ALNIS , P.
however, it cannot handle more complex multi-way or multi-level Authenticated join processing in outsourced databases. In SIGMOD (2009).
joins. Additionally, one may have to build multiple MHTs for a ta- [26] Z HOU , W., M AO , Y., L OO , B. T., AND A BADI , M. Unified Declarative
ble (each for a different join key) when using AIM on joins that are Platform for Secure Networked Information Systems. In ICDE (2009). See also
conducted with different join keys. http://netdb.cis.upenn.edu/ds2/.

6

Towards secure cloud data management

More Related Content

What's hot (18)

Viewers also liked (6)

Similar to Towards secure cloud data management (20)

More from ambitlick (20)

Recently uploaded (20)

Towards secure cloud data management