26-09-2016, 12:57 PM
1456201425-OnlineSignatureVerificationonMobileDevices.pdf (Size: 3.21 MB / Downloads: 30)
Abstract— This paper studies online signature verification on
touch interface-based mobile devices. A simple and effective
method for signature verification is developed. An online signature
is represented with a discriminative feature vector derived
from attributes of several histograms that can be computed in
linear time. The resulting signature template is compact and
requires constant space. The algorithm was first tested on the
well-known MCYT-100 and SUSIG data sets. The results show
that the performance of the proposed technique is comparable
and often superior to state-of-the-art algorithms despite its
simplicity and efficiency. In order to test the proposed method
on finger drawn signatures on touch devices, a data set was
collected from an uncontrolled environment and over multiple
sessions. Experimental results on this data set confirm the
effectiveness of the proposed algorithm in mobile settings. The
results demonstrate the problem of within-user variation of
signatures across multiple sessions and the effectiveness of cross
session training strategies to alleviate these problems.
INTRODUCTION
AHANDWRITTEN signature is a socially and legally
accepted biometric trait for authenticating an individual.
Typically, there are two types of handwritten signature verification
systems: off-line and online systems. In an off-line system,
just an image of the user’s signature is acquired without additional
attributes, whereas, in an online system, a sequence of
x-y coordinates of the user’s signature, along with associated
attributes like pressure, time, etc., are also acquired. As a
result, an online signature verification system usually achieves
better accuracy than an off-line system [1].
The increasing number of personal computing devices that
come equipped with a touch sensitive interface and the diffi-
culty of entering a password on such devices [2] have led to an
interest in developing alternative authentication mechanisms
on them [3], [4]. In this context, an online signature is a
plausible candidate given the familiarity users have with the
concept of using a signature for the purpose of authentication.
There has been much work on online signature verification
systems [5]–[10]. However, none of this has been directed
to the context of authentication on mobile devices. Previous work has addressed online signatures acquired from traditional
digitizers in a controlled environment. These are different from
those acquired from mobile devices in dynamic environments.
First, on mobile devices, a user performs his signatures in
various contexts, i.e., sitting or standing, mobile or immobile,
and holding a device at different angles and orientations.
Secondly, availability of computational resources may differ
from one signature instance to another and it could result
in greater variation of input resolution when compared to
that of stand-alone acquisition devices. Last, signatures on
mobile devices are often drawn using a finger instead of a
stylus resulting in less precise signals. An example of finger
drawn signatures acquired from mobile devices is depicted in
Figure 1.
Consequently, verification performance derived from traditional
datasets, collected using stylus-based devices in a
controlled environment, may not carry over to online signature
verification on mobile device setting [5]–[8]. In addition, other
characteristics of the system, i.e., a template aging [11] and
effectiveness of cross-sessions training, may be different when
signatures are obtained from a mobile device.
This paper proposes an online signature verification algorithm
that is suitable to deploy on mobile devices. It is a
computationally and space efficient algorithm for enrolling and
verifying signatures. In addition, a signature template is stored
in an irreversible form thereby providing privacy protection
to an original online signature. The proposed method was
evaluated on public datasets as well as new dataset collected
in in uncontrolled setting from user owned mobile devices.
The verification performance obtained is promising. The key
contributions made by this paper are as follows:
1) A method to extract a model-free non-invertible feature
set from an online signature is proposed. The
feature set comprises of sets of histograms that capture distributions of attributes generated from raw signature
data sequences and their combinations. By evaluating
the proposed method on public datasets, its verification
performance is superior to several state of the art
algorithms.
2) A new dataset was collected from 180 users in a mobile
device verification environment. The signatures in this
dataset were drawn with a fingertip, in an uncontrolled
setting on user owned iOS devices and over six separate
sessions with intervals ranging from 12 to 96 hours.
3) By applying the proposed method on the above dataset,
the following aspects of online signature verification on
mobile devices were investigated:
• impact of template aging on online signatures,
• effectiveness of using cross-session samples, or
samples from multiple sessions, to train a classifier,
and
• security of the system against random forgery, or
zero-effort attack, and its comparison to that of
4-digit PIN.
The rest of this paper is organized as follows. Section II
presents a process of deriving a set of histograms from an
online signature, gives details of the proposed online signature
verification system, and analyzes its complexity. Section III
provides experimental results on public datasets. In section IV,
a method and apparatus for collecting a new dataset as
well as experimental results and analysis on this dataset is
presented. Lastly, in section V, conclusions and future work
are discussed.
A. Previous Work
Typically, online signature verification techniques can be
classified into two approaches, namely, function-based and
feature-based [5]. The former refers to an approach where
the matching process is done using, directly or indirectly, the
original time series data points of a signature. The latter refers
to an approach where the matching process is done using
descriptive features of a signature. Examples of well-known
function-based approaches include Dynamic Time Warping
Algorithm (DTW) [6], [7], [12], and Hidden Markov Models
(HMM) [8].
A function-based system typically yields better verification
performance than a feature-based system [13]. However,
during the matching process, a dynamic construction of the
original signature is revealed resulting in a potential privacy
problem if the matching has to be done remotely. Furthermore,
the system is generally more complex and slower than featurebased
systems [6]. Even worse, when a template protection
approach is applied in order to provide biometric privacy
and/or security, verification performance often deteriorates
significantly. For instance, Maiorana et al [14] have proposed a
convolution scheme to protect the original signature sequence
of a user, that can be directly applied to any function based
approach. The idea is to split the original input sequence
into W subsequences. Each subsequence may have a different
length based on random parameters. This technique
has been applied with HMM and DTW based verification systems [14]–[16]. In each case, verification rates were lower
when compared to using the original versions of the signatures.
With a feature-based system, an online signature is represented
by a feature vector. Therefore, the original biometric
sample need not be stored or transmitted. Further, many known
algorithms [17], [18] can be used to derive cryptographic
keys from feature vectors. However, the main challenge for
a feature-based approach is to derive a descriptive set of
features that can be used to effectively and efficiently verify
an online signature [5], [6], [12]. In the literature, there are
many proposals to derive a set of features from an online
signature. In 2005, Fierrez-Aguilar et al [19] proposed a set
of 100 features, such as total duration of the signature, number
of pen ups, sign changes of (dx/dt) and (dy/dt), etc. to
represent a signature and applied a feature selection method
to rank the proposed features. Based on this 100 feature
set, Nanni [9] proposed a multi-matcher method to verify
an online signature. In addition, Guru and Prakash [10]
derived a symbolic representation of an online signature and
introduced the concept of writer independent threshold in
order to improve verification accuracy. Recently, Argones et
al [20] have proposed a set of HMM model features from
a universal background model. The best reported verification
performance obtained by their system is promising. However,
the system extracts 4800 features from tuning 16 different
HMM models, which is a computationally expensive task.
Moreover, the universal background model is trained from a
pool of 2500 genuine and forged signatures from 50 users on
the same device specification, where a user-specific classifier
is trained from 10 signatures. These make it less feasible to be
employed for mobile device authentication, where the embedded
sensors are different from model to model. In addition, the
HMM-based method is not robust to the well known templateaging
problem resulting in significant deterioration of verifi-
cation performance when verifying samples and enrollment
samples are from different sessions [11].
II. ONLINE SIGNATURE VERIFICATION ALGORITHM
As illustrated in Figure 2, the proposed system comprises
of three main components: a feature extractor, a template
generator, and a matcher. First, an online signature is processed
by the feature extractor in order to compute a set of histograms
from which a feature vector is derived. Then, the template
generator constructs a user-specific template using the feature
sets derived from multiple enrolled signatures. This template
is later used by the matcher to verify a test signature. The rest
of this section describes these three components in detail and
analyzes system complexity. (For more details, please refer
to [21] for the earlier version of this work.)
A. Feature Extractor
In the proposed system, an online signature is represented
by a set of histograms. These histogram features are designed
to capture essential attributes of the signature as well as
relationships between these attributes. It should be noted that
histograms are widely used as a feature set to capture attribute
statistics in many recognition tasks.