Privacy preserving detection of sensitive data exposure

1092 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 10, NO. 5, MAY 2015
Privacy-Preserving Detection of
Sensitive Data Exposure
Xiaokui Shu, Danfeng Yao, Member, IEEE, and Elisa Bertino, Fellow, IEEE
Abstract—Statistics from security firms, research institutions
and government organizations show that the number of data-leak
instances have grown rapidly in recent years. Among various
data-leak cases, human mistakes are one of the main causes
of data loss. There exist solutions detecting inadvertent sensitive
data leaks caused by human mistakes and to provide alerts
for organizations. A common approach is to screen content
in storage and transmission for exposed sensitive information.
Such an approach usually requires the detection operation to
be conducted in secrecy. However, this secrecy requirement is
challenging to satisfy in practice, as detection servers may be
compromised or outsourced. In this paper, we present a privacy-
preserving data-leak detection (DLD) solution to solve the issue
where a special set of sensitive data digests is used in detection.
The advantage of our method is that it enables the data owner to
safely delegate the detection operation to a semihonest provider
without revealing the sensitive data to the provider. We describe
how Internet service providers can offer their customers DLD as
an add-on service with strong privacy guarantees. The evaluation
results show that our method can support accurate detection
with very small number of false alarms under various data-leak
scenarios.
Index Terms—Data leak, network security, privacy, collection
intersection.
I. INTRODUCTION
ACCORDING to a report from Risk Based
Security (RBS) [2], the number of leaked sensitive
data records has increased dramatically during the last
few years, i.e., from 412 million in 2012 to 822 million
in 2013. Deliberately planned attacks, inadvertent leaks
(e.g., forwarding confidential emails to unclassified email
accounts), and human mistakes (e.g., assigning the wrong
privilege) lead to most of the data-leak incidents [3].
Detecting and preventing data leaks requires a set of
complementary solutions, which may include data-leak
detection [4], [5], data confinement [6]–[8], stealthy malware
detection [9], [10], and policy enforcement [11].
Network data-leak detection (DLD) typically performs deep
packet inspection (DPI) and searches for any occurrences
Manuscript received June 9, 2014; revised October 31, 2014 and
January 14, 2015; accepted January 22, 2015. Date of publication February 10,
2015; date of current version March 7, 2015. This work was supported by the
Security and Software Engineering Research Center, an NSF sponsored multi-
university Industry/University Cooperative Research Center. The associate
editor coordinating the review of this manuscript and approving it for
publication was Prof. Mauro Barni.
X. Shu and D. Yao are with the Department of Computer Science, Virginia
Tech, Blacksburg, VA, 24061 USA (e-mail: subx@vt.edu; danfeng@vt.edu).
E. Bertino is with the Department of Computer Science, Purdue University,
West Lafayette, IN 47907 USA (e-mail: bertino@cs.purdue.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIFS.2015.2398363
of sensitive data patterns. DPI is a technique to analyze
payloads of IP/TCP packets for inspecting application layer
data, e.g., HTTP header/content. Alerts are triggered when the
amount of sensitive data found in traffic passes a threshold.
The detection system can be deployed on a router or integrated
into existing network intrusion detection systems (NIDS).
Straightforward realizations of data-leak detection require
the plaintext sensitive data. However, this requirement is
undesirable, as it may threaten the confidentiality of the
sensitive information. If a detection system is compromised,
then it may expose the plaintext sensitive data (in memory).
In addition, the data owner may need to outsource the
data-leak detection to providers, but may be unwilling to reveal
the plaintext sensitive data to them. Therefore, one needs
new data-leak detection solutions that allow the providers
to scan content for leaks without learning the sensitive
information.
In this paper, we propose a data-leak detection solution
which can be outsourced and be deployed in a semi-
honest detection environment. We design, implement, and
evaluate our fuzzy fingerprint technique that enhances data
privacy during data-leak detection operations. Our approach
is based on a fast and practical one-way computation on the
sensitive data (SSN records, classified documents, sensitive
emails, etc.). It enables the data owner to securely delegate the
content-inspection task to DLD providers without exposing the
sensitive data. Using our detection method, the DLD provider,
who is modeled as an honest-but-curious (aka semi-honest)
adversary, can only gain limited knowledge about the sen-
sitive data from either the released digests, or the content
being inspected. Using our techniques, an Internet service
provider (ISP) can perform detection on its customers’ traffic
securely and provide data-leak detection as an add-on service
for its customers. In another scenario, individuals can mark
their own sensitive data and ask the administrator of their local
network to detect data leaks for them.
In our detection procedure, the data owner computes a
special set of digests or fingerprints from the sensitive data
and then discloses only a small amount of them to the
DLD provider. The DLD provider computes fingerprints from
network traffic and identifies potential leaks in them.
To prevent the DLD provider from gathering exact knowl-
edge about the sensitive data, the collection of potential
leaks is composed of real leaks and noises. It is the data
owner, who post-processes the potential leaks sent back by
the DLD provider and determines whether there is any real
data leak.
1556-6013 © 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted,
but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
www.redpel.com+917620593389

SHU et al.: PRIVACY-PRESERVING DETECTION OF SENSITIVE DATA EXPOSURE 1093
In this paper, we present details of our solution and provide
extensive experimental evidences and theoretical analyses to
demonstrate the feasibility and effectiveness of our approach.
Our contributions are summarized as follows.
1) We describe a privacy-preserving data-leak detection
model for preventing inadvertent data leak in network
traffic. Our model supports detection operation dele-
gation and ISPs can provide data-leak detection as an
add-on service to their customers using our model.
We design, implement, and evaluate an efficient
technique, fuzzy fingerprint, for privacy-preserving
data-leak detection. Fuzzy fingerprints are special sensi-
tive data digests prepared by the data owner for release
to the DLD provider.
2) We implement our detection system and perform exten-
sive experimental evaluation on 2.6 GB Enron dataset,
Internet surfing traffic of 20 users, and also 5 simulated
real-world data-leak scenarios to measure its privacy
guarantee, detection rate and efficiency. Our results indi-
cate high accuracy achieved by our underlying scheme
with very low false positive rate. Our results also show
that the detection accuracy does not degrade much
when only partial (sampled) sensitive-data digests are
used. In addition, we give an empirical analysis of our
fuzzification as well as of the fairness of fingerprint
partial disclosure.
II. MODEL AND OVERVIEW
We abstract the privacy-preserving data-leak detection
problem with a threat model, a security goal and a privacy
goal. First we describe the two most important players in
our abstract model: the organization (i.e., data owner) and the
data-leak detection (DLD) provider.
• Organization owns the sensitive data and authorizes the
DLD provider to inspect the network traffic from the
organizational networks for anomalies, namely inadver-
tent data leak. However, the organization does not want
to directly reveal the sensitive data to the provider.
• DLD provider inspects the network traffic for potential
data leaks. The inspection can be performed offline with-
out causing any real-time delay in routing the packets.
However, the DLD provider may attempt to gain knowl-
edge about the sensitive data.
We describe the security and privacy goals in
Section II-A and Section II-B.
A. Security Goal and Threat Model
We categorize three causes for sensitive data to appear on
the outbound traffic of an organization, including the legitimate
data use by the employees.
• Case I Inadvertent data leak: The sensitive data
is accidentally leaked in the outbound traffic by a
legitimate user. This paper focuses on detecting this
type of accidental data leaks over supervised network
channels. Inadvertent data leak may be due to human
errors such as forgetting to use encryption, carelessly for-
warding an internal email and attachments to outsiders,
or due to application flaws (such as described in [12]).
A supervised network channel could be an unencrypted
channel or an encrypted channel where the content in it
can be extracted and checked by an authority. Such a
channel is widely used for advanced NIDS where MITM
(man-in-the-middle) SSL sessions are established instead
of normal SSL sessions [13].
• Case II Malicious data leak: A rogue insider or a piece of
stealthy software may steal sensitive personal or organiza-
tional data from a host. Because the malicious adversary
can use strong private encryption, steganography or covert
channels to disable content-based traffic inspection, this
type of leaks is out of the scope of our network-based
solution. Host-based defenses (such as detecting the
infection onset [14]) need to be deployed instead.
• Case III Legitimate and intended data transfer: The
sensitive data is sent by a legitimate user intended for
legitimate purposes. In this paper, we assume that the data
owner is aware of legitimate data transfers and permits
such transfers. So the data owner can tell whether a piece
of sensitive data in the network traffic is a leak using
legitimate data transfer policies.
The security goal in this paper is to detect Case I leaks, that
is inadvertent data leaks over supervised network channels.
In other words, we aim to discover sensitive data appearance in
network traffic over supervised network channels. We assume
that: i) plaintext data in supervised network channels can
be extracted for inspection; ii) the data owner is aware of
legitimate data transfers (Case III); and iii) whenever sensitive
data is found in network traffic, the data owner can decide
whether or not it is a data leak. Network-based security
approaches are ineffective against data leaks caused by mal-
ware or rogue insiders as in Case II, because the intruder may
use strong encryption when transmitting the data, and both the
encryption algorithm and the key could be unknown to the
DLD provider.
B. Privacy Goal and Threat Model
To prevent the DLD provider from gaining knowledge of
sensitive data during the detection process, we need to set up a
privacy goal that is complementary to the security goal above.
We model the DLD provider as a semi-honest adversary,
who follows our protocol to carry out the operations, but
may attempt to gain knowledge about the sensitive data of
the data owner. Our privacy goal is defined as follows. The
DLD provider is given digests of sensitive data from the data
owner and the content of network traffic to be examined. The
DLD provider should not find out the exact value of a piece of
sensitive data with a probability greater than 1
K , where K is an
integer representing the number of all possible sensitive-data
candidates that can be inferred by the DLD provider.
We present a privacy-preserving DLD model with a new
fuzzy fingerprint mechanism to improve the data protection
against semi-honest DLD provider. We generate digests of
sensitive data through a one-way function, and then hide
the sensitive values among other non-sensitive values via
fuzzification. The privacy guarantee of such an approach

Fig. 1. Our Privacy-preserving Data-Leak Detection Model.
is much higher than 1
K when there is no leak in traffic,
because the adversary’s inference can only be gained through
brute-force guesses.
The traffic content is accessible by the DLD provider
in plaintext. Therefore, in the event of a data leak, the
DLD provider may learn sensitive information from the traffic,
which is inevitable for all deep packet inspection approaches.
Our solution confines the amount of maximal information
learned during the detection and provides quantitative
guarantee for data privacy.
C. Overview of Privacy-Enhancing DLD
Our privacy-preserving data-leak detection method supports
practical data-leak detection as a service and minimizes the
knowledge that a DLD provider may gain during the process.
Fig. 1 lists the six operations executed by the data owner and
the DLD provider in our protocol. They include PREPROCESS
run by the data owner to prepare the digests of sensitive
data, RELEASE for the data owner to send the digests to the
DLD provider, MONITOR and DETECT for the DLD provider
to collect outgoing traffic of the organization, compute digests
of traffic content, and identify potential leaks, REPORT for
the DLD provider to return data-leak alerts to the data owner
where there may be false positives (i.e., false alarms), and
POSTPROCESS for the data owner to pinpoint true data-leak
instances. Details are presented in the next section.
The protocol is based on strategically computing data
similarity, specifically the quantitative similarity between the
sensitive information and the observed network traffic. High
similarity indicates potential data leak. For data-leak detection,
the ability to tolerate a certain degree of data transformation
in traffic is important. We refer to this property as noise
tolerance. Our key idea for fast and noise-tolerant comparison
is the design and use of a set of local features that are
representatives of local data patterns, e.g., when byte b2
appears in the sensitive data, it is usually surrounded by bytes
b1 and b3 forming a local pattern b1, b2, b3. Local features
preserve data patterns even when modifications (insertion,
deletion, and substitution) are made to parts of the data. For
example, if a byte b4 is inserted after b3, the local pattern
b1, b2, b3 is retained though the global pattern (e.g., a hash
of the entire document) is destroyed. To achieve the privacy
goal, the data owner generates a special type of digests,
which we call fuzzy fingerprints. Intuitively, the purpose of
fuzzy fingerprints is to hide the true sensitive data in a crowd.
It prevents the DLD provider from learning its exact value.
We describe the technical details next.
III. FUZZY FINGERPRINT METHOD AND PROTOCOL
We describe technical details of our fuzzy fingerprint mech-
anism in this section.
A. Shingles and Fingerprints
The DLD provider obtains digests of sensitive data from
the data owner. The data owner uses a sliding window and
Rabin fingerprint algorithm [15] to generate short and hard-
to-reverse (i.e., one-way) digests through the fast polynomial
modulus operation. The sliding window generates small frag-
ments of the processed data (sensitive data or network traffic),
which preserves the local features of the data and provides
the noise tolerance property. Rabin fingerprints are computed
as polynomial modulus operations, and can be implemented
with fast XOR, shift, and table look-up operations. The
Rabin fingerprint algorithm has a unique min-wise inde-
pendence property [16], which supports fast random finger-
prints selection (in uniform distribution) for partial fingerprints
disclosure.
The shingle-and-fingerprint process is defined as follows.
A sliding window is used to generate q-grams on an input
binary string first. The fingerprints of q-grams are then
computed.
A shingle (q-gram) is a fixed-size sequence of contiguous
bytes. For example, the 3-gram shingle set of string
abcdefgh consists of six elements {abc, bcd, cde, def,
efg, fgh}. Local feature preservation is accomplished
through the use of shingles. Therefore, our approach can
tolerate sensitive data modification to some extent,
e.g., inserted tags, small amount of character substitution,
and lightly reformatted data. The use of shingles for finding
duplicate web documents first appeared in [17] and [18].
The use of shingles alone does not satisfy the one-wayness
requirement. Rabin fingerprint is utilized to satisfy such
requirement after shingling. In fingerprinting, each shingle
is treated as a polynomial q(x). Each coefficient of q(x),
i.e., ci (0 < i < k), is one bit in the shingle. q(x) is mod
by a selected irreducible polynomial p(x). The process shown
in (1) maps a k-bit shingle into a p f -bit fingerprint f where
the degree of p(x) is p f + 1.
f = c1xk−1
+ c2xk−2
+ . . . + ck−1x + ck mod p(x) (1)
From the detection perspective, a straightforward method
is for the DLD provider to raise an alert if any sensitive
fingerprint matches the fingerprints from the traffic.1 However,
this approach has a privacy issue. If there is a data leak,
there is a match between two fingerprints from sensitive
data and network traffic. Then, the DLD provider learns the
corresponding shingle, as it knows the content of the packet.
Therefore, the central challenge is to prevent the DLD provider
1In reality, data-leak detection solutions usually utilize more complex statis-
tical models to raise alerts instead of alerting individual fingerprints. Statistical
approaches, e.g., packet sensitivity in Section V, eliminate accidental matches.

from learning the sensitive values even in data-leak scenarios,
while allowing the provider to carry out the traffic inspection.
We propose an efficient technique to address this problem.
The main idea is to relax the comparison criteria by strategi-
cally introducing matching instances on the DLD provider’s
side without increasing false alarms for the data owner.
Specifically, i) the data owner perturbs the sensitive-data
fingerprints before disclosing them to the DLD provider,
and ii) the DLD provider detects leaking by a range-based
comparison instead of the exact match. The range used in the
comparison is pre-defined by the data owner and correlates
to the perturbation procedure. We define the notions of fuzzy
length and fuzzy set next and then describe how they are used
in our detailed protocol in Section III-B.
Definition 1: Given a p f -bit-long fingerprint f , the fuzzy
length pd (pd < p f ) is the number of bits in f that may be
perturbed by the data owner.
Definition 2: Given a fuzzy length pd, and a collection of
fingerprints, the fuzzy set S f,pd of a fingerprint f is the set of
fingerprints in the collection whose values differ from f by at
most 2pd − 1.
In Definition 2, the size of the fuzzy set |S f,pd | is upper
bounded by 2pd , but the actual size may be smaller due to the
sparsity of the fingerprint space.
B. Operations in Our Protocol
1) PREPROCESS: This operation is run by the data owner
on each piece of sensitive data.
a) The data owner chooses four public parameters
(q, p(x), pd, M). q is the length of a shingle. p(x),
is an irreducible polynomial (degree of p f + 1)
used in Rabin fingerprint. Each fingerprint is p f -bit
long and the fuzzy length is pd. M is a bitmask,
which is p f -bit long and contains pd 0’s at random
positions. The positions of 1’s and 0’s in
M indicate the bits to preserve and to randomize
in the fuzzification, respectively.
b) The data owner computes S, which is the set of
all Rabin fingerprints of the piece of sensitive
data.
c) The data owner transforms each fingerprint
f ∈ S into a fuzzy fingerprint f ∗ with ran-
domized bits (specified by the mask M). The
procedure is described as follows: for each f ∈ S,
the data owner generates a random p f -bit binary
string ˙f , mask out the bits not randomized by
˙f = (NOT M) AND ˙f (1’s in M indicate positions
of bits not to randomize), and fuzzify f with
f ∗ = f XOR ˙f . The overall computation is
described in (2).
f ∗
= ((NOT M) AND ˙f ) XOR f (2)
All fuzzy fingerprints are collected and form the
output of this operation, the fuzzy fingerprint
set, S∗.
2) RELEASE: This operation is run by the data owner.
The fuzzy fingerprint set S∗ obtained by PREPROCESS
is released to the DLD provider for use in the
detection, along with the public parameters
(q, p(x), pd, M). The data owner keeps S for use
in the subsequent POSTPROCESS operation.
3) MONITOR: This operation is run by the DLD provider.
The DLD provider monitors the network traffic T from
the data owner’s organization. Each packet in T is
collected and the payload of it is sent to the next
operation as the network traffic (binary) string ˜T .
The payload of each packet is not the only choice to
define ˜T . A more sophisticated approach could identify
TCP flows and extract contents in a TCP session as ˜T.
Contents of other protocols can also be retrieved if
required by the detection metrics.
4) DETECT: This operation is run by the DLD provider on
each ˜T as follows.
a) The DLD provider first computes the set of Rabin
fingerprints of traffic content ˜T based on the public
parameters. The set is denoted as T.
b) The DLD provider tests whether each fingerprint
f ∈ T is also in S∗ using the fuzzy equivalence
test (3).
E( f , f ∗
) = NOT (M AND ( f XOR f ∗
)) (3)
E( f , f ∗) is either True or False. f XOR f ∗
gives the difference between f and f ∗.
M AND ( f XOR f ∗) filters the result leaving only
the interesting bits (preserved bits with 1’s in M).
Because XOR yields 0 for equivalent bits, NOT
is used to turn 0-bits into 1’s (and 1’s into 0’s).
The overall result from (3) is read as a boolean
indicating whether or not f is equivalent to a
fuzzy fingerprint f ∗ ∈ S∗.
(2) and (3) are designed in a pair, and M works the
same in both equations by masking out fuzzified
bits at same positions in each f , f ∗ and f . All
f with True values are record in a set ˆT.
c) The DLD provider aggregates the outputs from
the preceding step and raises alerts based on a
threshold. Our concrete aggregation formula is
given in section V.
5) REPORT: If DETECTION on ˜T yields an alert, the
DLD provider reports the set of detected candidate leak
instances ˆT to the data owner.
6) POSTPROCESS: After receiving ˆT, the data owner test
every f ∈ ˆT to see whether it is in S. A precise
likelihood of data leaking is computed at the data
owner’s, which we discuss more in section V.
In the protocol, because S f ∗,pd , the fuzzy set of f ∗, includes
the original fingerprint f , the true data leak can be detected
(i.e., true positive). Yet, due to the increased detection range,
multiple values in S f ∗,pd may trigger alerts. Because S f ∗,pd
is large for the given network flow, the DLD provider has a
low probability of pinpointing the sensitive data, which can
be bounded as shown in Section IV.
The DETECT operation can be performed between T and S∗
via set intersection test with a special equivalence test function

(e.g. Formula 5 in Section V as one realization). The advantage
of our method is that the additional matching instances intro-
duced by fuzzy fingerprints protect the sensitive data from the
DLD provider; yet they do not cause additional false alarms
for the data owner, as the data owner can quickly distinguish
true and false leak instances. Given the digest f of a piece
of sensitive data, a large collection T of traffic fingerprints,
and a positive integer K |T |, the data owner can choose
a fuzzy length pd such that there are at least K − 1 other
distinct digests in the fuzzy set of f , assuming that the shingles
corresponding to these K digests are equally likely to be
candidates for sensitive data and to appear in network traffic.
A tight fuzzy length (i.e., the smallest pd value satisfying the
privacy requirement) is important for efficient POSTPROCESS
operation. Due to the dynamic nature of network traffic,
pd needs to be estimated accordingly. There exists an obvious
tradeoff between privacy and detection efficiency – large fuzzy
set allows a fingerprint to hide among others and confuses
the DLD provider, yet this indistinguishability results in more
work in POSTPROCESS. We provide quantitative analysis on
fuzzy fingerprint including empirical results on different sizes
of fuzzy sets.
C. Extensions
1) Fingerprint Filter: We develop this extension to use
Bloom filter in the DETECT operation for efficient set inter-
section test. Bloom filter [19] is a well-known space-saving
data structure for performing set-membership test. It applies
multiple hash functions to each of the set elements and stores
the resulting values in a bit vector; to test whether a value v
belongs to the set, the filter checks each corresponding bit
mapped with each hash function. Bloom filter in combination
with Rabin fingerprint is referred to by us as the fingerprint
filter. We use Rabin fingerprints with variety of modulus’s
in fingerprint filter as the hash functions, and we perform
extensive experimental evaluation on both fingerprint filter and
bloom filter with MD5/SHA in Section V.
2) Partial Disclosure: Using the min-wise independent
property of Rabin fingerprint, the data owner can quickly
disclose partial fuzzy fingerprints to the DLD provider. The
purpose of partial disclosure is two-fold: i) to increase the
scalability of the comparison in the DETECT operation, and
ii) to reduce the exposure of data to the DLD provider
for privacy. The method of partial release of sensitive data
fingerprints is similar to the suppression technique in database
anonymization [20], [21].
This extension requires a good uniform distribution random
selection to avoid disclosure bias. The min-wise indepen-
dence feature of Rabin fingerprint guarantees that the minimal
fingerprint is coming from a (uniformly distributed) random
shingle. The property is also valid for a minimum set of
fingerprints and so the data owner can just select r smallest
elements in S∗ to perform partial disclosure. The r elements
are then sent to the DLD provider in RELEASE operation
instead of S∗. We implement the partial disclosure policy,
evaluate its influence on detection rate, and verify the min-wise
independence property of Rabin fingerprint in Section V.
IV. ANALYSIS AND DISCUSSION
We analyze the security and privacy guarantees provided by
our data-leak detection system, as well as discuss the sources
of possible false negatives – data leak cases being overlooked
and false positives – legitimate traffic misclassified as data
leak in the detection. We point out the limitations associated
with the proposed network-based DLD approaches.
A. Privacy Analysis
Our privacy goal is to prevent the DLD provider from
inferring the exact knowledge of all sensitive data, both the
outsourced sensitive data and the matched digests in network
traffic. We quantify the probability for the DLD provider to
infer the sensitive shingles as follows.
A polynomial-time adversary has no greater than 2
p f −pd
n
probability of correctly inferring a sensitive shingle, where
p f is the length of a fingerprint in bits, pd is the fuzzy
length, and n ∈ [2p f −pd , 2p f ] is the size of the set of traffic
fingerprints, assuming that the fingerprints of shingles are
uniformly distributed and are equally likely to be sensitive
and appear in the traffic.
We explain our quantification in two scenarios:
i) There is a match between a sensitive fuzzy finger-
print f ∗ (derived from the sensitive fingerprint f ) and
fingerprints from the network traffic. Because the size of
fuzzy set S f,pd is upper bounded by 2pd (Definition 2),
there could be at most 2pd (sensitive or non-sensitive)
fingerprints fuzzified into the identical f ∗. Given a
set (size n) of traffic fingerprints, the DLD provider
expects to find K fingerprints matched to f ∗
where K = n
2
p f × 2pd .
a) If f corresponds to a sensitive shingle leaked in the
traffic, i.e., f is within the K traffic fingerprints,
the likelihood of correctly pinpointing f from the
K fingerprints is 1
K , or 2
p f −pd
n . The likelihood is
fare because both sensitive data and network traffic
contain binary data. It is difficult to predict the
subspace of the sensitive data in the entire binary
space.
b) If the shingle form of f is not leaked in the traffic,
the DLD provider cannot use the K traffic finger-
prints, which match f ∗, to infer f . Alternatively,
the DLD provider needs to brute force f ∗ to get f ,
which is discussed as in the case of no match.
ii) There is no match between sensitive and traffic finger-
prints. The adversarial DLD provider needs to brute
force reverse the Rabin fingerprinting computation to
obtain the sensitive shingle. There are two difficulties
in reversing a fingerprint: i) Rabin fingerprint is a
one-way hash. ii) Multiple shingles can map to the
same fingerprint. It requires to searching the complete
set of possible shingles for a fingerprint and to identify
the sensitive one from the set. This brute-force attack
is difficult for a polynomial-time adversary, thus the
success probability is not included.
In summary, the DLD provider cannot decide whether
the alerts (traffic fingerprints matched f ∗) contain any leak

or not (case i.a or i.b). Even if it is known that there is a real
leak in the network traffic, the polynomial-time DLD provider
has no greater than 2
p f −pd
n probability of correctly pinpointing
a sensitive shingle (case i.a).
B. Alert Rate
We quantify the rate of alerts expected in the traffic for a
sensitive data entry (the fuzzified fingerprints set of a piece
of sensitive data) given the following values: the total number
of fuzzified sensitive fingerprints τ, the expected traffic fin-
gerprints set size n, fingerprint length p f , fuzzy length pd,
partial disclosure rate ps ∈ (0, 1], and the expected rate α,
which is the percentage of fingerprints in the sensitive data
entry that appear in the network traffic. The expected alert
rate R is presented in (4).
R =
αps Kτ
n
=
αpsτ
2p f −pd
(4)
R is used to derive threshold Sthres in the detection;
Sthres should be lower than R. The overhead of our privacy-
preserving approach over traditional fingerprinting data-leak
detection solutions is tightly related to R and Sthres, because
there is an extra POSTPROCESS step in our approach after the
DLD provider detects potential leaks. The less potential leaks
the DLD provider reports back to the data owner, the less
overhead is introduced by our privacy-preserving approach,
while the less privacy is achieved since K is small.
C. Collisions
Collisions may be due to where the legitimate traffic
happens to contain the partial sensitive-data fingerprints by
coincidence. The collision may increase with shorter shingles,
or smaller numbers of partial fingerprints, and may decrease
if additional features such as the order of fingerprints are
used for detection. A previous large-scale information-retrieval
study empirically demonstrated the low rate of this type
of collisions in Rabin fingerprint [18], which is a desir-
able property suggesting low unwanted false alarms in our
DLD setting. Collisions due to two distinct shingles generating
the same fingerprint are proved to be low [17] and are
negligible.
D. Space of Sensitive Data
The space of all text-based sensitive data may be smaller
than the space of all possible shingles. Yet, when including
non-ASCII sensitive data (text in UTF-8 or binaries), the space
of sensitive data can be significantly expanded. A restricted
space limits K and can expose the fuzzified fingerprint. For
instance, one may assume that a password has higher entropy
than normal English shingles, thus a fuzzy fingerprint of a
password rules out much space of S f,pd where normal English
lives. The full space with a variety of text encodings and
binaries ensures that the major space is still there shadowing
the fuzzy fingerprint.
E. Short Polynomial Modulus
A naive alternative to the fuzzy fingerprint mechanism is to
use a shorter polynomial modulus to compute Rabin finger-
prints (e.g., 16-bit instead of 32-bit). This approach increases
collisions for fuzzification purpose. However, one issue of this
approach is that true positive and false positives yield the same
fingerprint value due to collision, which prevents the data
owner from telling true positives apart from false positives.
In addition, our fuzzy fingerprint approach is more flexible
from the deployment perspective, as the data owner can adjust
and fine-tune the privacy and accuracy in the detection without
recomputing the fingerprints. In contrast, the precision is fixed
in the naive shorter polynomial approach unless fingerprints
are recomputed.
Limitations: we point out three major limitations of our
detection approach within our threat model.
F. Modified Data Leak
The underlying shingle scheme of our approach has limited
power to capture heavily modified data leaks. False negatives
(i.e., failure to detect data leak) may occur due to the data
modification (e.g., reformatting). The new shingles/fingerprints
may not resemble the original ones, and cannot be detected.
As a result, a packet may evade the detection. In our exper-
iments, we evaluate the impact of several types of data
transformation in real world scenarios. The modified data-leak
detection problem is a general problem for all comparison-
based data-leak detection solutions. More advanced content
comparison techniques than shingles/fingerprints are needed
to fully address the issue.
G. Dynamic Sensitive Data
For protecting dynamically changing data such as source
code or documents under constant development or keystroke
data, the digests need to be continuously updated for detection,
which may not be efficient or practical. We raise the issue of
how to efficiently detect dynamic data with a network-based
approach as an open problem to investigate by the community.
H. Selective Fragments Leak
The partial disclosure scheme may result in false negatives,
i.e., the leaked data may evade the detection because it is
not covered by the released fingerprints. This issue illustrates
the tradeoff among detection accuracy, privacy guarantee and
detection efficiency. Fortunately, it is expensive for an attacker
to escape the detection with partial disclosure. On one hand,
Rabin fingerprint guarantees that every fingerprint has the
same probability to be selected and released through its min-
wise independence property. Deliberately choosing unreleased
segments from sensitive data is not easy. On the other hand,
even figuring out which fingerprints are not released, one needs
leaking inconsecutive bytes to bypass the detection. It usually
makes no sense to leak inconsecutive bytes from sensitive
data. Some format, e.g., binary, may be destroyed through the
leaking.

V. EXPERIMENTAL EVALUATION
We implement our fuzzy fingerprint framework in Python,
including packet collection, shingling, Rabin fingerprinting,
as well as partial disclosure and fingerprint filter extensions.
Our implementation of Rabin fingerprint is based on cyclic
redundancy code (CRC). We use the padding scheme men-
tioned in [22] to handle small inputs. In all experiments,
the shingles are in 8-byte, and the fingerprints are in 32-bit
(33-bit irreducible polynomials in Rabin fingerprint). We set
up a networking environment in VirtualBox, and make a
scenario where the sensitive data is leaked from a local
network to the Internet. Multiple users’ hosts (Windows 7)
are put in the local network, which connect to the Internet
via a gateway (Fedora). Multiple servers (HTTP, FTP, etc.)
and an attacker-controlled host are put on the Internet side.
The gateway dumps the network traffic and sends it to a
DLD server/provider (Linux). Using the sensitive-data finger-
prints defined by the users in the local network, the DLD server
performs off-line data-leak detection. The speed aspect of
privacy-preserving data-leak detection is another topic and we
study it in [23].
In our prototype system, the DLD server detects the
sensitive data within each packet on basis of a stateless filtering
system. We define the sensitivity of a packet in (5), which is
used by the DLD provider in DETECTION. It indicates the
likelihood of a packet containing sensitive data.
Spacket =
|
pd
¨S∗ ∩
pd
T|
min(|S∗|, |T|)
×
|S∗|
|¨S∗|
(5)
T is the set of all fingerprints extracted in a packet. S∗
is the set of all sensitive fuzzy fingerprints. For each piece
of sensitive data, the data owner computes S∗ and reveals a
partial set ¨S∗ (¨S∗ ⊆ S∗) to the DLD provider. The operator
t
indicates right shifting every fingerprint in a set by t bits,
which is the implementation of a simple mask M in our
protocol (Section III-B). |S∗|/|¨S∗| estimates the leaking level
of S∗ according to the revealed and tested partial set ¨S∗. When
too few fuzzy fingerprints are revealed, e.g., 10%, the samples
may not sufficiently describe the leaking characteristic of the
traffic, and the estimation can be imprecise. For each packet,
the DLD server computes Spacket (Spacket ∈ [0, 1]). If it is
higher than a threshold Sthres ∈ (0, 1), T is reported back
to the data owner, and the data owner uses (6) to determine
whether it is a real leak in POSTPROCESS.
Spacket =
|S ∩ T|
min(|S|, |T|)
(6)
The difference between (5) operated by the DLD provider
and (6) by the data owner is that the original fingerprints S are
used in (6) instead of the sampled and fuzzified set ¨S∗ in (5),
so the data owner can pinpoint the exact leaks.
The use of Spacket and Sthres for detection is important
because individual shingles or fingerprints are not accurate
features to represent an entire piece of sensitive data.
Sensitive data can share strings with non-sensitive data,
e.g., formatting strings, which results in occasionally reported
sensitive fingerprints. Spacket is an accumulated score and
Sthres filters out packets with occasionally discovered sensitive
fingerprints.
The evaluation goal is to answer the following questions:
1) Can our solution accurately detect sensitive data leak
in the traffic with low false positives (false alarms) and
high true positives (real leaks)?
2) Does using partial sensitive-data fingerprints reduce the
detection accuracy in our system?
3) What is the performance advantage of our fingerprint
filter over traditional Bloom filter with SHA-1?
4) How to choose a proper fuzzy length and make a balance
between the privacy need and the number of alerts?
In the following subsection, we experimentally addressed
and answered all the questions. For the first three questions, we
present results based on the Spacket value calculated in (6). The
first and second questions are answered in Section V-A. The
third question is discussed in Section V-B. The last question
is designed to understand the properties of fuzzification and
partial disclosure, and it is addressed in Section V-C.
A. Accuracy Evaluation
We evaluate the detection accuracy in simple and complex
leaking scenarios. First we test the detection rate and false pos-
itive rate in three simple experiments where the sensitive data
is leaked in its original form or not leaked. Then we present
accuracy evaluation on more complex leaking experiments to
reproduce various real-world leaking detection scenarios.
1) Simple Leaking Scenarios: We test our prototype without
partial disclosure in simple leaking scenarios, i.e., ¨S∗ = S∗.
We generate 20,000 personal financial records as the sensitive
data and store them in a text file. The data contains person
name, social security number, credit card number, credit card
expiration date, and credit card CVV.
To evaluate the accuracy of our strategy, we perform three
separate experiments using the same sensitive dataset:
Exp.1 True leak A user leaks the entire set of sensitive data
via FTP by uploading it to a remote FTP server.
Exp.2 No leak The non-related outbound HTTP traffic of
20 users is captured (30 minutes per user), and given
to the DLD server to analyze. No sensitive data (i.e.,
zero true positive) should be confirmed.
Exp.3 No leak The Enron dataset (2.6 GB data, 150 users’
517,424 emails) as a virtual network traffic is given
to the DLD server to analyze. Each virtual network
packet created is based on an email in the dataset.
No sensitive data (i.e., zero true positive) should be
confirmed by the data owner.
The detection results are shown in Table I. Among the three
experiments, the first one is designed to infer true positive rate.
We manually check each packet and the DLD server detects
all 651 real sensitive packets (all of them have sensitivity
values greater than 0.9). The sensitivity value is less than one,
because the high-layer headers (e.g., HTTP) in a packet are not
sensitive. The next two experiments are designed to estimate
the false positive rate. We found that none of the packets has
a sensitivity value greater than 0.05. The results indicate that
our design performs as expected on plaintext.

Fig. 2. Detection accuracy comparison in terms of (a) the averaged sensitivity and (b) the number of detected sensitive packets. X-axis is the partial disclosure
rate, or the percentage of sensitive-data fingerprints revealed to the DLD server and used in the detection. [out] indicates outbound traffic only, while [all]
means both outbound and inbound traffic captured and analyzed.
TABLE I
MEAN AND STANDARD DEVIATIONS OF THE SENSITIVITY PER PACKET IN
THREE SEPARATE EXPERIMENTS. FOR EXP.1, THE HIGHER SENSITIVITY,
THE BETTER; FOR THE OTHER TWO (NEGATIVE CONTROL),
THE LOWER SENSITIVITY, THE BETTER
2) Complex Leaking Scenarios: The data owner may reveal
a subset of sensitive data’s fingerprints to the DLD server for
detection, as opposed to the entire set. We are particularly
interested in measuring the percentage of revealed fingerprints
that can be detected in the traffic, assuming that fingerprints
are equally likely to be leaked.2 We reproduce four real-world
scenarios where data leaks are caused by human users or
software applications.
Exp.4 Web leak: a user posts sensitive data on wiki
(MediaWiki) and blog (WordPress) pages.
Exp.5 Backdoor leak: a program (Glacier) on the user’s
machine (Windows 7) leaks sensitive data.
Exp.6 Browser leak: a malicious Firefox extension
FFsniFF records the information in sensitive web
forms, and emails the data to the attacker.
Exp.7 Keylogging leak: a keylogger EZRecKb exports
intercepted keystroke values on a user’s host.3 It con-
nects to a SMTP server on the Internet side and sends
its log of keystrokes periodically.
In these four experiments, the source file of TCP/IP
page on Wikipedia (24KB in text) is used as the sensitive
data. The detection is performed at various partial disclo-
sure rate. The subset of the sensitive fingerprints is selected
randomly. The sensitivity threshold is Sthres = 0.05.
2Given the subset independence property, sensitive-data’s fingerprints are
equally likely to be selected for detection.
3EZRecKb records every key stroke and replaces the function keys with
labels, such as [left shift].
Fig. 2 shows the comparison of performance across various
size of fingerprints used in the detection, in terms of the
averaged sensitivity per packet in (a) and the number of
detected sensitive packets in (b). These accuracy values reflect
results of the POSTPROCESS operation.
The results show that the use of partial sensitive-data
fingerprints does not much degrade the detection rate
compared to the use of full sets of sensitive-data fingerprints.
However, extreme small sampling rates, e.g., 10%, may not
provide sufficient numbers of fingerprints to describe the
leaking characteristic of the traffic. The packet sensitivity esti-
mation (|S|/|¨S| in (6)) magnifies the signal (the real sensitivity
of a packet) as well as the noise produced by fingerprint
sampling. The precision could be affected and drops, e.g.,
6 packets with 10% vs. 3 packets with 100% for Keylogger
in Fig. 2 (b).
In Fig. 2 (a), the sensitivities of experiments vary due to
different levels of modification by the leaking programs, which
makes it difficult to perform detection. WordPress substi-
tutes spaces with +’s when sending the HTTP POST request.
EZRecKb inserts function-key as labels into the original text.
Typing typos and corrections also bring in modification to the
original sensitive data. In Fig. 2 (b), [all] results contain both
outbound and inbound traffic and double the real number of
sensitive packets in Blog and Wiki scenarios due to HTML
fetching and displaying of the submitted data.
B. Runtime Comparison
Our fingerprint filter implementation is based on the
Bloom filter library in Python (Pybloom). We compare
the runtime of Bloom filter provided by standard Pybloom
(with dynamically selected hash function from MD5, SHA-1,
SHA-256, SHA-384 and SHA-512) and that of fingerprint
filter with Rabin fingerprint. For Bloom filters and fingerprint
filters, we test their performance with 2, 6, and 10 hash
functions. We inspect 100 packets with random content against
10 pieces sensitive data at various lengths – there are a
total of 1,625,600 fingerprints generated from the traffic and
76,160 pieces of fingerprints from the sensitive data.

Fig. 3. The overhead of using the filters to detect data leaks. The detection
time is averaged from 100 packets against all 10 pieces of sensitive data.
(a) Detection Time. (b) Filter Building Time.
We present the time for detection per packet in Fig. 3 (a).
It shows that fingerprint filters run faster than Bloom filters,
which is expected as Rabin fingerprint is easier to compute
than MD5/SHA. The gap is not significant due to the fact that
Python uses a virtualization architecture. We have the core
hash computations implemented in Python C/C++ extension,
but the remaining control flow and function call statements are
in pure Python. The performance difference between Rabin
fingerprint and MD5/SHA is largely masked by the runtime
overhead spent on non-hash related operations.
In Fig. 3 (a), the number of hash functions used in Bloom
filters does not significantly impact their runtime, because
only one hash function is operated in most cases for Bloom
filters. Pybloom automatically chooses SHA-256 for Bloom
filter with 6 hash functions and SHA-384 for Bloom filter
with 10 hash functions. One hash is sufficient to distinguish
32-bits fingerprints. MD5 is automatically chosen for the
Bloom filter with 2 hash functions, which gives more collisions
and the second hash could be involved. We speculate this is the
reason why Bloom filter with 2 hashes is slower than Bloom
filters with 6 or 10 hashes. All of our fingerprint filters use
32-bit Rabin fingerprint functions. The small output space
Fig. 4. The observed and expected sizes of fuzzy sets per fingerprint (32-bit)
in Brown Corpus dataset (in blue) and network traffic (in red) with different
fuzzy lengths.
requires more than one hash for a membership test, so there is
more significant overhead when a fingerprint filter is equipped
with more hashes (6 vs. 2 and 10 vs. 6).
The filter construction time is shown in Fig. 3 (b). It shares
similar characteristics with the detection time. Filters with
more hash functions require more time to initialize, because
every hash function need to be computed. The construc-
tion of fingerprint filters, especially assigning the irreducible
polynomials p(x) for each Rabin fingerprint, is written in
pure Python, which is significantly slower than SHA-256 and
SHA-384 encapsulated using Python C/C++ extension.
C. Sizes of Fuzzy Sets vs. Fuzzy Length
The size of fuzzy set corresponds to the K value in our
definition of privacy goal. The higher K is, the more difficult
it is for a DLD provider to infer the original sensitive data
using our fuzzy fingerprinting mechanism – the fingerprint of
the sensitive data hides among its neighboring fingerprints.
We empirically evaluate the average size of the fuzzy set
associated with a given fuzzy length with both Brown Corpus
(text) and real-world network traffic (text & binary).
• Brown Corpus: The Brown University Standard
Corpus of Present-Day American English [24]. It contains
500 samples of English text across 15 genres, and there
are 1,014,312 words in total.
• Network traffic: 500MB Internet traffic dump collected
by us on a single host. It includes a variety of network
traffic: multimedia Internet surfing (images, video, etc.),
binary downloading, software and system updates, user
profile synchronization, etc.
We aim to show the trend of how the fuzzy-set sizes
changes with the fuzzy length, which can be used to select the
optimal fuzzy length used in the algorithm. We compute 32-bit
fingerprints from the datasets, and then count the number of
neighbors for each fingerprint. Fig. 4 shows the estimated and
observed sizes of fuzzy sets for fuzzy lengths in the range
of [14, 27] for 218,652 and 189,878 fingerprints generated
from the Brown Corpus dataset and the network traffic dataset.

The figure shows that the empirical results observed are very
close with the expected values of the fuzzy set sizes computed
based on our analysis in Section IV. This close fit also
indicates the uniform distribution of the fingerprints.
The fuzzy set is small when the fuzzy length is small,
which is due to the sparsity nature of Rabin fingerprints.
Given an estimated composition of traffic content, the data
owner can use the result of this experiment to determine
the optimal fuzzy length. In the datasets evaluated in the
experiments, for fuzzy length of 26 and 27 bits, the K values
are above 1,500 and 3,000, respectively. Because the data
owner can defuzzify in POSTPROCESS very quickly, the false
positives can be sifted out by the data owner. We also find
that for a fixed fuzzy length the distribution of fuzzy-set sizes
follows a Gaussian distribution. Different datasets may have
different K size characteristics. We demonstrate the feasibility
of estimating the fuzzy set sizes, which illustrates how fuzzy
fingerprintings can be used to realize a privacy goal.
1) Summary: Our detection rates in terms of the number
of sensitive packets found do not decrease much with the
decreasing size of disclosed fingerprint sets in Fig. 2, even
when only 10% of the sensitive-data fingerprints are used for
detection. Our experiments evaluate several noisy conditions
such as noise insertion – MediaWiki-based leak scenario,
and data substitution – for the keylogger- and WordPress-
based leak scenarios. Our results indicate that our fingerprint
filter can tolerate these three types of noises in the traffic
to some degree. Our approach works well especially in the
case where consecutive data blocks are leaked (i.e., local data
features are preserved). When the noises spread across the data
and destroy the local features (e.g., replacing every space with
another character), the detection rate decreases as expected.
The use of shorter shingles mitigates the problem, but it may
increase false positives. How to improve the noise tolerance
property in those conditions remains an open problem. Our
fuzzy fingerprint mechanism supports the detection of data-
leak at various sizes and granularities. We study the fuzzy set
size and also verify the min-wise independence property of
Rabin fingerprint, which are the building blocks of our fuzzy
fingerprint method.
VI. RELATED WORK
There have been several advances in understanding the
privacy needs [25] or the privacy requirement of security
applications [26]. In this paper, we identify the privacy needs
in an outsourced data-leak detection service and provide
a systematic solution to enable privacy-preserving
DLD services.
Shingle with Rabin fingerprint [15] was used previously
for identifying similar spam messages in a collaborative
setting [27], as well as collaborative worm containment [28],
virus scan [29], and fragment detection [30].
In comparison, we tackle the unique data-leak detection
problem in an outsourced setting where the DLD provider
is not fully trusted. Such privacy requirement does not exist
in above models, e.g., the virus signatures are non-sensitive in
the virus-scan paradigm [29]. We propose the fuzzy fingerprint
approach to meet the special privacy requirement and present
the first systematic solution to privacy-preserving data-leak
detection with convincing results.
Most data-leak detection products offered by the indus-
try, e.g., Symantec DLP [31], Global Velocity [32], Identity
Finder [4], do not have the privacy-preserving feature and
cannot be outsourced. GoCloudDLP [33] is a little different,
which allows its customers to outsource the detection to a fully
honest DLD provider. Our fuzzy fingerprint method differs
from these solutions and enables its adopter to provide data-
leak detection as a service. The customer or data owner does
not need to fully trust the DLD provider using our approach.
Bloom filter [19] is a space-saving data structure for set
membership test, and it is used in network security from
network layer [34] to application layer [35]. The fuzzy Bloom
filter invented in [36] constructs a special Bloom filter that
probabilistically sets the corresponding filter bits to 1’s.
Although it is designed to support a resource-sufficient routing
scheme, it is a potential privacy-preserving technique. We do
not invent a variant of Bloom filter for our fuzzy fingerprint,
and our fuzzification process is separate from membership
test. The advantage of separating fingerprint fuzzification from
membership test is that it is flexible to test whether the
fingerprint is sensitive with or without fuzzification.
Besides fingerprint-based detection, other approaches can be
applied to data-leak detection. If the sensitive data size is small
and the patterns of all sensitive data are enumerable, string
matching [37], [38] in network intrusion detection system
(such as SNORT [39] or Bro [40]) can be used to detect
data leaks. Privacy-preserving keyword search [41] or fuzzy
keyword search [42] provide string matching approaches in
semi-honest environments, but keywords usually do not cover
enough sensitive data segments for data-leak detection.
Anomaly detection in network traffic can be used to detect
data leaks. [5] detects any substantial increase in the amount
of new information in the traffic, and entropy analysis is used
in [43]. We present a signature-based model to detect data
leaks and focus on the design that can be outsourced, thus the
two approaches are different.
Another category of approaches for data-leak detection is
tracing and enforcing the sensitive data flows. The approaches
include data flow and taint analysis [6], legal flow mark-
ing [44], and file-descriptor sharing enforcement [8]. These
approaches are different from ours because they do not aim
to provide an remote service. However, pure network-based
solution cannot handle maliciously encrypted traffic [45], and
these methods are complementary to our approach in detecting
different forms (e.g., encrypted) of data leaks.
Besides our fuzzy fingerprint solution for data-leak detec-
tion, there are other privacy-preserving techniques invented for
specific processes, e.g., DNA matching [46], or for general
purpose use, e.g., secure multi-party computation (SMC).
Similar to string matching methods discussed above, [46] uses
anonymous automata to perform comparison. SMC [47] is a
cryptographic mechanism, which supports a wide range of
fundamental arithmetic, set, and string operations as
well as complex functions such as knapsack computa-
tion [48], automated trouble-shooting [49], network event
statistics [50], [51], private information retrieval [52],

genomic computation [53], private database query [54], private
join operations [55], and distributed data mining [56]. The
provable privacy guarantees offered by SMC comes at a cost
in terms of computational complexity and realization difficulty.
The advantage of our approach is its concision and efficiency.
VII. CONCLUSIONS AND FUTURE WORK
We proposed fuzzy fingerprint, a privacy-preserving
data-leak detection model and present its realization. Using
special digests, the exposure of the sensitive data is kept to a
minimum during the detection. We have conducted extensive
experiments to validate the accuracy, privacy, and efficiency of
our solutions. For future work, we plan to focus on designing a
host-assisted mechanism for the complete data-leak detection
for large-scale organizations.
ACKNOWLEDGMENT
The authors would like to thank Michael Wolfe for his
insightful comments and feedback.
REFERENCES
[1] X. Shu and D. Yao, “Data leak detection as a service,” in Proc. 8th Int.
Conf. Secur. Privacy Commun. Netw., 2012, pp. 222–240.
[2] Risk Based Security. (Feb. 2014). Data Breach Quick-
View: An Executive’s Guide to 2013 Data Breach Trends.
[Online]. Available: https://www.riskbasedsecurity.com/reports/2013-
DataBreachQuickView.pdf, accessed Oct. 2014.
[3] Ponemon Institute. (May 2013). 2013 Cost of Data Breach Study: Global
Analysis. [Online]. Available: https://www4.symantec.com/mktginfo/
whitepaper/053013_GL_NA_WP_Ponemon-2013-Cost-of-a-Data-
Breach-Report_daiNA_cta72382.pdf, accessed Oct. 2014.
[4] Identity Finder. Discover Sensitive Data Prevent Breaches DLP Data
Loss Prevention. [Online]. Available: http://www.identityfinder.com/,
accessed Oct. 2014.
[5] K. Borders and A. Prakash, “Quantifying information leaks in outbound
web traffic,” in Proc. 30th IEEE Symp. Secur. Privacy, May 2009,
pp. 129–140.
[6] H. Yin, D. Song, M. Egele, C. Kruegel, and E. Kirda, “Panorama:
Capturing system-wide information flow for malware detection and
analysis,” in Proc. 14th ACM Conf. Comput. Commun. Secur., 2007,
pp. 116–127.
[7] K. Borders, E. V. Weele, B. Lau, and A. Prakash, “Protecting confidential
data on personal computers with storage capsules,” in Proc. 18th
USENIX Secur. Symp., 2009, pp. 367–382.
[8] A. Nadkarni and W. Enck, “Preventing accidental data disclosure in
modern operating systems,” in Proc. 20th ACM Conf. Comput. Commun.
Secur., 2013, pp. 1029–1042.
[9] A. Kapravelos, Y. Shoshitaishvili, M. Cova, C. Kruegel, and G. Vigna,
“Revolver: An automated approach to the detection of evasiveweb-based
malware,” in Proc. 22nd USENIX Secur. Symp., 2013, pp. 637–652.
[10] X. Jiang, X. Wang, and D. Xu, “Stealthy malware detection and monitor-
ing through VMM-based ‘out-of-the-box’ semantic view reconstruction,”
ACM Trans. Inf. Syst. Secur., vol. 13, no. 2, 2010, p. 12.
[11] G. Karjoth and M. Schunter, “A privacy policy model for enterprises,”
in Proc. 15th IEEE Comput. Secur. Found. Workshop, Jun. 2002,
pp. 271–281.
[12] J. Jung, A. Sheth, B. Greenstein, D. Wetherall, G. Maganis, and
T. Kohno, “Privacy oracle: A system for finding application leaks
with black box differential testing,” in Proc. 15th ACM Conf. Comput.
Commun. Secur., 2008, pp. 279–288.
[13] Y. Jang, S. P. Chung, B. D. Payne, and W. Lee, “Gyrus: A framework for
user-intent monitoring of text-based networked applications,” in Proc.
23rd USENIX Secur. Symp., 2014, pp. 79–93.
[14] K. Xu, D. Yao, Q. Ma, and A. Crowell, “Detecting infection onset
with behavior-based policies,” in Proc. 5th Int. Conf. Netw. Syst. Secur.,
Sep. 2011, pp. 57–64.
[15] M. O. Rabin, “Fingerprinting by random polynomials,” Dept. Math.,
Hebrew Univ. Jerusalem, Jerusalem, Israel, Tech. Rep. TR-15-81, 1981.
[16] A. Z. Broder, M. Charikar, A. M. Frieze, and M. Mitzenmacher,
“Min-wise independent permutations,” J. Comput. Syst. Sci., vol. 60,
no. 3, pp. 630–659, 2000.
[17] A. Z. Broder, “Some applications of Rabin’s fingerprinting method,” in
Sequences II. New York, NY, USA: Springer-Verlag, 1993, pp. 143–152.
[18] A. Z. Broder, “Identifying and filtering near-duplicate documents,” in
Proc. 11th Annu. Symp. Combinat. Pattern Matching, 2000, pp. 1–10.
[19] A. Broder and M. Mitzenmacher, “Network applications of bloom filters:
A survey,” Internet Math., vol. 1, no. 4, pp. 485–509, 2004.
[20] G. Aggarwal et al., “Anonymizing tables,” in Proc. 10th Int. Conf.
Database Theory, 2005, pp. 246–258.
[21] R. Chen, B. C. M. Fung, N. Mohammed, B. C. Desai, and K. Wang,
“Privacy-preserving trajectory data publishing by local suppression,” Inf.
Sci., vol. 231, pp. 83–97, May 2013.
[22] M. O. Rabin, “Digitalized signatures and public-key functions as
intractable as factorization,” Massachusetts Inst. Technol., Cambridge,
MA, USA, Tech. Rep. MIT/LCS/TR-212, 1979.
[23] F. Liu, X. Shu, D. Yao, and A. R. Butt, “Privacy-preserving scanning
of big content for sensitive data exposure with MapReduce,” in Proc.
ACM CODASPY, 2015.
[24] W. N. Francis and H. Kucera, “Brown corpus manual,” 1979.
[25] J. Kleinberg, C. H. Papadimitriou, and P. Raghavan, “On the value of
private information,” in Proc. 8th Conf. Theoretical Aspects Rationality
Knowl., 2001, pp. 249–257.
[26] S. Xu, “Collaborative attack vs. collaborative defense,” in Collaborative
Computing: Networking, Applications and Worksharing (Lecture Notes
of the Institute for Computer Sciences, Social Informatics and Telecom-
munications Engineering), vol. 10. Berlin, Germany: Springer-Verlag,
2009, pp. 217–228.
[27] K. Li, Z. Zhong, and L. Ramaswamy, “Privacy-aware collaborative spam
filtering,” IEEE Trans. Parallel Distrib. Syst., vol. 20, no. 5, pp. 725–739,
May 2009.
[28] M. Cai, K. Hwang, Y.-K. Kwok, S. Song, and Y. Chen, “Collaborative
Internet worm containment,” IEEE Security Privacy, vol. 3, no. 3,
pp. 25–33, May 2005.
[29] F. Hao, M. Kodialam, T. V. Lakshman, and H. Zhang, “Fast payload-
based flow estimation for traffic monitoring and network security,” in
Proc. ACM Symp. Archit. Netw. Commun. Syst., Oct. 2005, pp. 211–220.
[30] L. Ramaswamy, A. Iyengar, L. Liu, and F. Douglis, “Automatic detection
of fragments in dynamically generated web pages,” in Proc. 13th Int.
Conf. World Wide Web, 2004, pp. 443–454.
[31] Symantec. Data Loss Prevention (DLP) Software. [Online]. Available:
http://www.symantec.com/data-loss-prevention/, accessed Oct. 2014.
[32] Global Velocity Inc. Cloud Data Security From the Inside Out. [Online].
Available: http://www.globalvelocity.com/, accessed Oct. 2014.
[33] GTB Technologies Inc. SaaS Content Control in the Cloud.
[Online]. Available: http://www.gtbtechnologies.com/en/solutions/dlp_as
_a_service, accessed Oct. 2014.
[34] S. Geravand and M. Ahmadi, “Bloom filter applications in network
security: A state-of-the-art survey,” Comput. Netw., vol. 57, no. 18,
pp. 4047–4064, Dec. 2013.
[35] B. Wang, S. Yu, W. Lou, and Y. T. Hou, “Privacy-preserving multi-
keyword fuzzy search over encrypted data in the cloud,” in Proc. 33th
IEEE Conf. Comput. Commun., Apr./May 2014, pp. 2112–2120.
[36] C. P. Mayer, “Bloom filters and overlays for routing in pocket switched
networks,” in Proc. 5th Int. Student Workshop Emerg. Netw. Experim.
Technol., 2009, pp. 43–44.
[37] A. V. Aho and M. J. Corasick, “Efficient string matching: An aid to
bibliographic search,” Commun. ACM, vol. 18, no. 6, pp. 333–340, 1975.
[38] P.-C. Lin, Y.-D. Lin, Y.-C. Lai, and T.-H. Lee, “Using string matching
for deep packet inspection,” IEEE Comput., vol. 41, no. 4, pp. 23–28,
Apr. 2008.
[39] M. Roesch et al., “Snort—Lightweight intrusion detection for networks,”
in Proc. 13th USENIX Conf. Syst. Admin., 1999, pp. 229–238.
[40] V. Paxson, “Bro: A system for detecting network intruders in real-time,”
Comput. Netw., vol. 31, nos. 23–24, pp. 2435–2463, 1999.
[41] S. Ananthi, M. Sadish Sendil, and S. Karthik, “Privacy preserving
keyword search over encrypted cloud data,” in Advances in Comput-
ing and Communications (Communications in Computer and Infor-
mation Science), vol. 190. Berlin, Germany: Springer-Verlag, 2011,
pp. 480–487.
[42] J. Li, Q. Wang, C. Wang, N. Cao, K. Ren, and W. Lou, “Fuzzy keyword
search over encrypted data in cloud computing,” in Proc. 29th IEEE
Conf. Comput. Commun., Mar. 2010, pp. 1–5.

[43] T. W. Fawcett, “ExFILD: A tool for the detection of data
exfiltration using entropy and encryption characteristics of network
traffic,” M.S. thesis, Dept. Elect. Comput. Eng., Univ. Delaware, Newark,
DE, USA, 2010.
[44] J. Croft and M. Caesar, “Towards practical avoidance of information
leakage in enterprise networks,” in Proc. 6th USENIX Conf. Hot Topics
Secur. (HotSec), 2011, p. 7.
[45] V. Varadharajan, “Internet filtering—Issues and challenges,” IEEE Secu-
rity Privacy, vol. 8, no. 4, pp. 62–65, Jul./Aug. 2010.
[46] J. R. Troncoso-Pastoriza, S. Katzenbeisser, and M. Celik, “Privacy
preserving error resilient dna searching through oblivious automata,”
in Proc. 14th ACM Conf. Comput. Commun. Secur., 2007, pp. 519–528.
[47] A. C.-C. Yao, “How to generate and exchange secrets,” in Proc. 27th
Annu. Symp. Found. Comput. Sci., 1986, pp. 162–167.
[48] D. Yao, K. B. Frikken, M. J. Atallah, and R. Tamassia, “Private
information: To reveal or not to reveal,” ACM Trans. Inf. Syst. Secur.,
vol. 12, no. 1, 2008, Art. ID 6.
[49] Q. Huang, D. Jao, and H. J. Wang, “Applications of secure electronic
voting to automated privacy-preserving troubleshooting,” in Proc. 12th
ACM Conf. Comput. Commun. Secur., 2005, pp. 68–80.
[50] P. Williams and R. Sion, “Usable PIR,” in Proc. 13th Netw. Distrib. Syst.
Secur. Symp., 2008.
[51] M. Burkhart, M. Strasser, D. Many, and X. Dimitropoulos, “SEPIA:
Privacy-preserving aggregation of multi-domain network events and
statistics,” in Proc. 19th USENIX Conf. Secur. Symp., 2010, p. 15.
[52] X. Yi, R. Paulet, and E. Bertino, Private Information Retrieval (Syn-
thesis Lectures on Information Security, Privacy, & Trust). San Rafael,
CA, USA: Morgan & Claypool Pub., 2013.
[53] S. Jha, L. Kruger, and V. Shmatikov, “Towards practical privacy for
genomic computation,” in Proc. 29th IEEE Symp. Secur. Privacy,
May 2008, pp. 216–230.
[54] X. Yi, M. G. Kaosar, R. Paulet, and E. Bertino, “Single-database private
information retrieval from fully homomorphic encryption,” IEEE Trans.
Knowl. Data Eng., vol. 25, no. 5, pp. 1125–1134, May 2013.
[55] B. Carbunar and R. Sion, “Joining privately on outsourced data,”
in Secure Data Management (Lecture Notes in Computer Science),
vol. 6358. Berlin, Germany: Springer-Verlag, 2010, pp. 70–86.
[56] G. Jagannathan and R. N. Wright, “Privacy-preserving distributed
k-means clustering over arbitrarily partitioned data,” in Proc. 11th ACM
SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2005, pp. 593–599.
Xiaokui Shu is currently pursuing the Ph.D. degree
in computer science with the Virginia Polytech-
nic Institute and State University (Virginia Tech),
Blacksburg, VA, USA, and his research is in system
and network security. He received the bachelor’s
degree in information security from the University
of Science and Technology of China (USTC), Hefei,
China, in 2010. Being the top student in the class, he
graduated with the Guo Moruo Award, and received
the SIMIT Chinese Academy of Sciences Scholar-
ship in 2008. He succeeded at his first real-world
penetration test at USTC and received the first prize in the Virginia Tech
Inaugural Cyber Security Summit Competition in 2011.
Danfeng Yao is an Associate Professor with the
Department of Computer Science, Virginia Polytech-
nic Institute and State University, Blacksburg, VA,
USA. She received the Ph.D. degree in computer sci-
ence from Brown University, Providence, RI, USA,
in 2007. Her recent research has been focused on
network and system security and assurance, includ-
ing anomaly detection in networked systems and
programs, system integrity, malware program analy-
sis, botnet detection, and data exfiltration prevention.
She was a recipient of the NSF CAREER Award
for her work on human-behavior driven malware detection in 2010, and
most recently the ARO Young Investigator Award for her semantic reasoning
for mission-oriented security work in 2014. She was also a recipient of
the Outstanding New Assistant Professor Award from the Virginia Tech
College of Engineering in 2012. She received several best paper awards
(ICICS’06, CollaborateCom’09, and ICNP’12). She also received the award
for Technological Innovation from Brown University in 2006. She held a
U.S. patent for her anomaly detection technologies. She is an Associate Editor
of the IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING.
She serves as a Program Committee Member in numerous computer security
conferences, including ACM CCS. She has authored over 60 peer-reviewed
publications.
Elisa Bertino is currently a Professor with the
Department of Computer Science, Purdue Univer-
sity, West Lafayette, IN, USA, where she is also
the Research Director of the Center for Education
and Research in Information Assurance and Secu-
rity. Her research interests cover many areas in the
fields of information security and database systems.
Her work combines both theoretical and practical
aspects, addressing applications on a number of
domains, such as medicine and humanities. She is
the Editor-in-Chief of the IEEE TRANSACTIONS ON
DEPENDABLE AND SECURE COMPUTING. She serves on the Editorial Boards
of several security journals, such as the ACM Transactions on Information
and System Security and the IEEE Security and Privacy Magazine. She is
a fellow of the Association for Computing Machinery. She was a recipient
of the 2002 IEEE Computer Society Technical Achievement Award for
outstanding contributions to database systems and database security and
advanced data management systems, the 2005 Tsutomu Kanai Award by the
IEEE Computer Society for pioneering and innovative research contributions
to secure distributed systems, and most recently the 2014 ACM SIGSAC
Outstanding Contributions Award for her seminal research contributions and
outstanding leadership to data security and privacy for the past 25 years.

Privacy preserving detection of sensitive data exposure

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Privacy preserving detection of sensitive data exposure (20)

More from redpel dot com (20)

Recently uploaded (20)

Privacy preserving detection of sensitive data exposure