6472 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO.

10, OCTOBER 2022

SynSig2Vec: Forgery-Free Learning of Dynamic

Signature Representations by Sigma
Lognormal-Based Synthesis and 1D CNN
Songxuan Lai , Lianwen Jin , Yecheng Zhu, Zhe Li, and Luojun Lin

Abstract—Handwritten signature verification is a challenging task because signatures of a writer may be skillfully imitated by a forger. As
skilled forgeries are generally difficult to acquire for training, in this paper, we propose a deep learning-based dynamic signature verification
framework, SynSig2Vec, to address the skilled forgery attack without training with any skilled forgeries. Specifically, SynSig2Vec consists of a
novel learning-by-synthesis method for training and a 1D convolutional neural network model, called Sig2Vec, for signature representation
extraction. The learning-by-synthesis method first applies the Sigma Lognormal model to synthesize signatures with different distortion levels
for genuine template signatures, and then learns to rank these synthesized samples in a learnable representation space based on average
precision optimization. The representation space is achieved by the proposed Sig2Vec model, which is designed to extract fixed-length
representations from dynamic signatures of arbitrary lengths. Through this training method, the Sig2Vec model can extract extremely effective
signature representations for verification. Our SynSig2Vec framework requires only genuine signatures for training, yet achieves state-of-the-
art performance on the largest dynamic signature database to date, DeepSignDB, in both skilled forgery and random forgery scenarios.
Source codes of SynSig2Vec will be available at https://github.com/LaiSongxuan/SynSig2Vec.

Index Terms—Dynamic signature verification and synthesis, Sigma Lognormal, average precision optimization, Sig2Vec

1 INTRODUCTION Typically, a signature verification system faces two types

of forgery attacks, namely random forgeries and skilled
VER the centuries, the handwritten signature has been a
O well-established method for personal authentication
under various scenarios, such as contracts, last wills and
forgeries. Random forgeries are produced by a forger who
has no knowledge regarding the authentic author’s name or
signature, and can be easily detected owing to their distinct
financial transactions. Because of its enduring and wide-
geometric and kinematic properties from the genuine signa-
spread use in history, the handwritten signature is recog-
tures. By contrast, skilled forgeries are produced by a mas-
nized in today’s e-society as an important biometric trait
ter forger who observes the signature of a writer and
facing both technological and cultural considerations. Dur-
attempts to mimic it with a similar intra-class variability.
ing the past 40 years, research interest in automatic signature
Because human’s handwriting naturally exhibits a large
verification has grown steadily, and a number of comprehen-
intra-class variability, the skilled forgery attack remains an
sive survey papers have well summarized the state-of-the-
open problem in signature research [4].
art in this research field as of 2019 [1], [2], [3], [4]. In general,
During the past several years, following the broad suc-
a signature verification system either processes static signa-
cess of deep learning [5], learning deep representations for
tures, which are scanned images of signatures written on
dynamic signatures has considerably reduced the verifica-
paper documents, or dynamic signatures, which are time
tion error against skilled forgeries [6], [7], [8], [9], [10], [11],
functions (e.g., trajectory, velocity and pressure) recorded by
[12]. More recently, as the database scales up [13], the verifi-
electronic devices during the signing process. In this paper,
cation error has been further decreased [14]. However, these
we focus on dynamic signature verification.
methods require skilled forgeries for training to achieve a
good performance. One should know that, as a biometric

Songxuan Lai, Yecheng Zhu, and Zhe Li are with the School of Electronic trait and a special type of private data, handwritten signa-
and Information Engineering, South China University of Technology,
Guangzhou 510641, China. E-mail: [email protected], {461438818,
tures are non-trivial to collect; skilled forgeries, which
eelizhe}@qq.com. require conscientious practice by the forgers, are even more

Lianwen Jin is with the School of Electronic and Information Engineering, difficult to acquire. Although a large-scale western signature
South China University of Technology, Guangzhou 510641, China, and also database with skilled forgeries is now available [13], toward
with the Guangdong Artificial Intelligence and Digital Economy Laboratory,
Pazhou Lab, Guangzhou 510335, China. E-mail: [email protected]. dealing with different scripts and different signature acquisi-

Luojun Lin is with the School of Mathematics and Computer Science, Fuzhou tion conditions, it is still preferable to effectively train a signa-
University, Fuzhou 350108, China. E-mail: [email protected]. ture verification system using only genuine signatures.
Manuscript received 17 June 2020; revised 26 Mar. 2021; accepted 1 June 2021. Besides the above-mentioned issue, we also notice that it
Date of publication 8 June 2021; date of current version 9 Sept. 2022. has not been investigated whether and how convolutional
(Corresponding author: Lianwen Jin.)
Recommended for acceptance by C. V. Jawahar. neural networks (CNNs) can be applied to obtain a discrim-
Digital Object Identifier no. 10.1109/TPAMI.2021.3087619 inative and holistic representation, i.e., feature vector, from
0162-8828 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6473

Fig. 1. Our SynSig2Vec framework for learning dynamic signature representations. SynSig2Vec consists of a novel learning-by-synthesis method for
training and the 1D CNN-based Sig2Vec model. Specifically, given a template signature, we extract its SL parameters and introduce perturbations of
two strength levels; two groups of distorted synthetic signatures can thus be reconstructed from the perturbed parameters (green and blue colors
indicate the dataflow of high and low perturbation synthesis respectively), and the Sig2Vec model learns to rank these synthesized samples with
supervision information from the perturbation process. Different colors in the SigVec model simply denote different feature channels. The ranking
branch optimizes the average precision of ranking, while the classification branch performs ordinary classification of signature contents.

time functions of a dynamic signature. Previous methods overfit the training set, possibly leading to severe performance
either use recurrent neural networks (RNNs) to achieve this degradation on the target domain data.
purpose [6], [9], [10], or instead learn a local feature sequence With regard to the second issue, a novel 1D CNN model
[7], [8], [11], [12], [14], generally with dynamic time warping is proposed to extract holistic representations from dynamic
(DTW) for sequence alignment to boost the performance [8], signatures of arbitrary lengths. We term our model as Sig2-
[11], [12], [14]. Both RNNs and DTW are time-consuming Vec (signature to vector), which vividly describes its func-
compared with CNNs. Therefore, it is desired to investigate a tionality. Specifically, inspired by multi-head attention [24],
proper CNN architecture to aggregate the time functions into a novel selective pooling (SP) module is designed to effec-
a holistic representation, which is also called signature tively pool output feature sequences of 1D convolutional
“parameters” in some studies [15], [16]. layers with learnable queries. Compared with traditional
In this paper, we propose novel techniques to address the average pooling, the SP can attend to different local struc-
above issues in deep learning-based dynamic signature ver- tures and aggregate their information into different sub-vec-
ification. For the first issue, under a scenario where no tors. With the proposed Sig2Vec, we can extract signature
skilled forgeries are provided as training data, we propose a representations that substantially defeat the benchmark
novel learning-by-synthesis method based on the kinematic method, i.e., running DTW on raw time functions.
theory of rapid human movements and its Sigma Lognor- Our framework is described in Fig. 1, and is termed as
mal (SL) model [17]. The SL model hypothesizes that the SynSig2Vec for short to highlight the two important afore-
velocity of a neuromuscular system can be modelled by a mentioned aspects. By training Sig2Vec with the learning-
vector summation of a number of lognormal functions, each by-synthesis method, we achieve a very competitive result
described by six parameters. Rooted in this model, we on the DeepSignDB database [13], compared with the state-
extract the underlying parameters of genuine signatures of-the-art TA-RNN model [14]. When trained with both
and synthesize new signatures by introducing perturbations genuine signatures and skilled forgeries, Sig2Vec performs
to the parameters. The level of parameter perturbation con- even better against skilled forgery attacks, demonstrating
trols the level of distortion to the template signature, hence its great capability in learning signature representations.
we can synthesize signatures with various distortion levels. The main contributions of this study are summarized as
Thereafter, we construct a novel 1D CNN to learn to rank follows: (1) based on the SL model, we devise a novel learn-
these synthesized signatures based on their cosine similari- ing-by-synthesis method that allows training deep dynamic
ties to the template signature in the feature space, and opti- signature verification systems without use of skilled forger-
mize the average precision (AP) of the ranking through the ies; (2) we introduce ranking-based AP optimization and
direct loss minimization framework [18]. Compared with demonstrate its effectiveness for dynamic signature repre-
other representation learning methods, AP optimization sentation learning; (3) we propose the Sig2Vec model to
can preserve and exploit fine-grained signature similarities extract discriminative and holistic representations from
in the ranking list and leads to a better verification accuracy. dynamic signatures of arbitrary lengths; (4) we conduct
The core of our learning-by-synthesis method is that the extensive experiments on the largest dynamic signature
supervision information comes from the prior SL parameter database to date, DeepSignDB, and achieve state-of-the-art
perturbation process. Previous methods only use the SL results under various evaluation scenarios.
model for signature generation [19], [20], [21] or duplication Major extensions in this paper over its conference version
[22], [23], whereas our method incorporates the SL model into [25] include: (1) we simultaneously synthesize pen-downs
a self-supervised learning framework. Furthermore, to the best and pen-ups, whereas in the conference paper, we only syn-
of our knowledge, we first demonstrate that synthesized thesized pen-downs and connected them with virtual pen-
dynamic signatures can be used to train deep learning-based ups; (2) a novel selective pooling module is introduced in
systems that can effectively verify real handwritten signatures. the Sig2Vec model to effectively pool feature sequences of
This outcome is important, as deep learning models generally arbitrary lengths, using multi-head attention with learnable
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6474 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

queries; (3) we propose other non-trivial modifications, such paper we propose a novel learning-by-synthesis method to
as feature fusion in the Sig2Vec model and new time func- address this issue.
tions as inputs; (4) we conduct extensive experiments and
an in-depth analysis on the DeepSignDB database as com- 2.2 Dynamic Signature Synthesis
pared to previous small-scale databases.
Dynamic signature synthesis is a promising method to com-
The rest of this paper is organized as follows. In Section 2,
bat the lack of data in dynamic signature verification. The
we review related work on dynamic signature verification
core of most synthesis methods is the kinematic theory of
and synthesis. In Section 3, we describe our framework in
rapid human movements and its SL model [17]. A synthetic
detail, including SL-based signature synthesis, AP optimi-
signature can either be fully synthesized [19], [20], [21] by
zation, the Sig2Vec model, and the signature verifier. In Sec-
trajectory generation and kinematic property manipulation,
tion 4, we present and analyse the experiment results on the
or be duplicated [22], [23] from a template signature accord-
DeepSignDB database. Finally, In Section 5, we conclude
ing to its SL parameters. Herein, we consider both fully
the paper with possible future research directions.
synthetic signatures and duplicated signatures to be syn-
thetic, as opposed to real handwritten signatures.
Galbally et al. [19] and Ferrer et al. [20], [21] showed that
2 RELATED WORK similar detection error trade-offs curves could be obtained
2.1 Dynamic Signature Verification for synthetic and real signature databases, as evidenced by
A recent survey paper [4] has comprehensively summarized several traditional verification systems. Visual Turing tests
the advancements of both dynamic and static signature ver- were also conducted to confirm the naturalness of the syn-
ifications in the last 10 years. During this period, although thesized signatures. However, they did not investigate
deep learning has became a de facto standard in pattern rec- whether synthesized signatures can be used to train deep
ognition and dominated most biometric systems [26], most learning-based systems that can effectively verify real hand-
state-of-the-art dynamic signature verification systems are written signatures. This question is important, because any
still based on traditional approaches, with DTW being the bias introduced in the synthesis process will undoubtedly
dominating method so far [4]. Because deep learning relies be overfitted by the networks, possibly leading to severe
heavily on training data, the research of dynamic signature performance degradation. Ahrabian and Babaali [6] trained
verification was encumbered by the lack of a large-scale and tested their deep learning models on fully synthetic sig-
database until the recent release of DeepSignDB [13]. natures, released by Ferrer et al. [20], and still did not
Nevertheless, in the absence of a large-scale database, a answer the above question. Indeed, their nearly perfect
number of researches still successfully proved the feasibility of results on the synthetic database seemed to be an evidence
applying deep learning models in dynamic signature verifica- that their model had overfitted the synthesis algorithm.
tion. These studies can be roughly divided into two categories. Diaz et al. [22] duplicated auxiliary signatures from one real
The first category learns local representations [7], [8], [11], [12], signature to enhance the performance of single reference
[14], generally with DTW for sequence alignment to boost the signature systems, while Ferrer at al. [23] tried to generate
performance. For example, Lai and Jin [8] proposed gated auto forgeries from genuine signatures to alleviate the drawback
regressive units, an RNN variant, to learn local feature sequen- of collecting skilled forgeries; nevertheless, neither of these
ces, and used DTW to measure the distance of the learned two approaches applied deep learning-based systems.
sequences. Wu et al. [11] and Tolosana et al. [14] used DTW to In this work, we give a positive answer to the above
align signature pairs in advance, and then input them into a question. Based on the SL model, we synthesize (or, more
CNN or RNN Siamese [27] network for similarity comparison. precisely, duplicate) signatures with different distortion lev-
The second category learns holistic signature representations els to the genuine template, and construct Sig2Vec to learn
[6], [9], [10], [28]. Ahrabian and Babaali [6] proposed stacked to rank signatures with lower distortion levels before those
LSTM [29] autoencoders to learn a fixed-length latent space, with higher distortion levels. Through this training method,
based upon which a Siamese network was further trained for very robust signature representations can be obtained for
verification. Park et al. [9] and Li et al. [10] divided a signature real handwritten signature verification.
into strokes and proposed to learn representations at both
stroke and signature levels, using a CNN-LSTM hybrid net- 3 METHODOLOGY OF SYNSIG2VEC
work and pure LSTMs respectively. Sekhar et al. [28] extracted 3.1 Learning-by-Synthesis
numerous global features manually, e.g., max velocity and
The proposed learning-by-synthesis method enables the
average pressure, and built a lightweight 1D CNN upon these
training of deep learning-based dynamic signature verifica-
features to derive a compact and discriminative feature vector.
tion systems without use of any skilled forgeries. It consists
Limited by the small database scale, previous methods have
of a SL-based signature synthesis algorithm and a ranking-
not fully exploited the potential of deep learning. For example,
based learning algorithm, which are described below.
they usually have a very parsimonious model. As the database
scales up [13], it is time to investigate more advanced deep
learning techniques in the field of dynamic signature verifica- 3.1.1 SL-Based Signature Synthesis
tion. In this paper, we propose a novel 1D CNN model, Sig2- The kinematic theory of rapid human movements [17], from
Vec, to extract holistic representations from time functions of which the SL model was developed, suggests that human
dynamic signatures. Another limitation in previous methods handwriting consists in controlling the pen-tip velocity with
is that they require training with skilled forgeries, and in this overlapped lognormal impulse responses, called strokes, as
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6475

difference is a non-zero residual trajectory ðx" ðtÞ; y" ðtÞÞ, that

contains no valid strokes according to the parameter extraction
algorithm. We consider the residual trajectory because it may
still contain useful information. By introducing perturbations
to the parameter matrix P and properly aligning the paramet-
ric and residual trajectories, new synthetic signatures can be
generated.1

Algorithm 1. Signature Synthesis Algorithm

Input: signature duration T , SL parameter matrix P , residual
trajectory ðx" ðtÞ; y" ðtÞÞ, pressure pðtÞ.
Output: a synthesized signature.
1: Sample RDi , Rt0;i , Rmi , Rsi , Rus;i , Rue;i .
2: Perturb SL parameters using Eqs. (6)-(13): P ! P^.
3: Start and end moments of the parametric trajectory:
4: Ts ¼ expð^ m1 Þ  expðm1 Þ þ Iðt0;1 < 0Þ  t0;1  Rt0;1 ,
5: Te ¼ expð^ m1 Þ  expðm1 Þ þ T  ð1 þ Rt0;1 Þ.
Fig. 2. The velocity profile of a typical human handwriting component 6: Recover the Rparametric trajectory from P^:
t PN
consists of lognormal impulse responses, called strokes. Based on these 7: x^p ðtÞ ¼ Ts i¼1 j~ vi ðt; P^i Þjcosðfi ðt; P^i ÞÞdt;
strokes’ parameters, the velocity and trajectory can be recovered. R t PN
8: y^p ðtÞ ¼ Ts i¼1 j~ vi ðt; P^i Þjsinðfi ðt; P^i ÞÞdt;
illustrated in Fig. 2. The magnitude and direction of the veloc- 9: where t 2 ½Ts ; Te . ⊳ Note that integral starts from Ts .
vi ðtÞ of the ith stroke is described as
ity profile ~ 10: Interpolate x" ðtÞ, y" ðtÞ, pðtÞ from time interval ½0; T  to
  ½0; Te  Ts : x" ðtÞ ! x^" ðtÞ, y" ðtÞ ! y^" ðtÞ, pðtÞ ! p^ðtÞ.
Di ðlnðt  t0;i Þ  mi Þ2 11: Align and add the parametric and residual trajectories:
vi ðt; Pi Þj ¼ pffiffiffiffiffiffi
j~ exp ;
2ps i ðt  t0;i Þ 2s 2i 12: x^ðtÞ ¼ x^p ðt þ Ts Þ þ x^" ðtÞ, y^ðtÞ ¼ y^p ðt þ Ts Þ þ y^" ðtÞ,
(1) 13: where t 2 ½0; Te  Ts .
Z 14: Return ð^ xðtÞ; y^ðtÞ; p^ðtÞÞ.
ue;i  us;i t
fi ðt; Pi Þ ¼ us;i þ j~
vi ðt; Pi Þjdt; (2)
Di 0 Specifically, Di , t0;i , mi and s i are perturbed as follows:
T
where Pi ¼ ½Di ; t0;i ; mi ; s i ; us;i ; ue;i  , Di is the amplitude, t0;i
^ i ¼ Di ð1 þ RD Þ;
D (6)
is the time occurrence, mi is the log time delay, s i is the log i

response time, and us;i and ue;i are the starting and ending
angles of the ith stroke respectively. The velocity of a com- t^0;i ¼ t0;i ð1 þ Rt0;i Þ; (7)
plete handwriting movement is considered as a vector sum-
mation of the individual stroke velocities
^ i ¼ mi ð1 þ Rmi Þ;
m (8)
X
N
vðtÞ ¼
~ vi ðt; Pi Þ;
~ (3)
i¼1 s^ i ¼ s i ð1 þ Rsi Þ; (9)

with N being the number of strokes. Thus, each stroke is while us;i , ue;i are perturbed as follows:
defined by six parameters in Pi , and a complete handwrit-
ing component is parameterized by P ¼ ½P1 ; P2 ; . . . ; PN . ^us;i ¼ us;i þ Ru ; (10)
s;i
Herein, one complete “component” refers to an entire signa-
ture trajectory, including both pen-ups and pen-downs. In ^ue;i ¼ ue;i þ Ru : (11)
e;i
the conference version [25], we considered each pen-down
as one component. However, joining pen-ups and pen- Here, RDi , Rt0;i , Rmi , Rsi , Rus;i , Rue;i are uniform random var-
downs together as one single component has been found to iables that decide the signature distortion level, and their
improve the verification accuracy, which is demonstrated in values are randomly set for each stroke. Furthermore, we
the experiments. reset t^0;i and m
^ i as follows:
Based on the SL parameter matrix P , a trajectory can be
recovered as follows: t^0;i ¼ t0;i ð1 þ Rt0;1 Þ; (12)

Z tX
N
xp ðtÞ ¼ j~
vi ðt; Pi Þjcosðfi ðt; Pi ÞÞdt; (4) 1. For extraction of SL parameters, each template signature is
0 i¼1 resampled at 200 Hz, following the practice in [31], [32], and filtered
Z tX
N
with a Butterworth low-pass filter with a cutoff frequency of 10 Hz.
After parameter extraction following the method in [31], the parametric
yp ðtÞ ¼ j~
vi ðt; Pi Þjsinðfi ðt; Pi ÞÞdt: (5) trajectory is subtracted from the non-filtered signature trajectory to
0 i¼1 obtain the residual trajectory, which mainly contains high frequency
information. Synthetic signatures are generated according to Algorithm
In practice, the parametric trajectory ðxp ðtÞ; yp ðtÞÞ is very close, 1, and down-sampled from 200 Hz to 100 Hz, which is the sampling
but not exactly equal, to the original trajectory ðxðtÞ; yðtÞÞ; the rate of most dynamic signature databases.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6476 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

TABLE 1
Configurations of the Random Variables That Decide the Signature Distortion Levels

Parameters Admissible Ranges Distortion Levels

G1 G2
Di Not applicable* RDi  U½0:025; 0:025 RDi  U½0:065; 0:040 [ ½0:040; 0:065
t0;i lt0 = 0.0825, ht0 = 0.0850. Rt0;i  U½0:25lt0 ; 0:25ht0  Rt0;i  U½0:65lt0 ; 0:40lt0  [ ½0:40ht0 ; 0:65ht0 
lm h lm lm h h
mi lm = 0.3950, hm = 0.3775. Rmi  U½0:25 1:5 ; 0:25 1:5m  Rmi  U½0:65 1:5 ; 0:40 1:5  [ ½0:40 1:5m ; 0:65 1:5m 
si ls = 0.3250, hs = 0.2875. Rsi  U½0:25ls ; 0:25hs  Rsi  U½0:65ls ; 0:40ls  [ ½0:40ls ; 0:65hs 
us;i Not used Rus;i  U½0:10; 0:10 Rus;i  U½0:20; 0:20
ue;i Not used Rue;i  U½0:10; 0:10 Rue;i  U½0:20; 0:20

Parameter Di does not affect the overall shape in Bhattacharya et al.’s synthesis setting [30]: they used component-wise parameter perturbation as opposed to
our stroke-wise perturbation.

mi Þ  expðmi Þ ¼ expð^
expð^ m1 Þ  expðm1 Þ; (13) Specifically, a neural network can be viewed as a composite
scoring function F ðx; y; wÞ, which depends on the input x 2
for i ¼ 2; . . . ; N. Specific constraints are applied to t^0;i and X , the output y 2 Y, and some parameters w. The theorem
^ i , so that t0;i ! t^0;i controls the contraction or expansion of
m (cited here for completeness) states that:
the generated signature, while mi ! m ^ i decides a same shift
Therom 1. When given a finite set Y, a scoring function
direction and distance of the velocity peaks, according to
F ðx; y; wÞ, a data distribution, as well as a task-loss Lðy; y^Þ,
Eq. (1). Thereby, we can properly align the parametric tra-
then, under some mild regularity conditions, the direct loss gra-
jectory with the residual trajectory, and add them together
dient has the following form:
to synthesize a new signature, as described in Algorithm 1.
Besides the trajectory, a pressure time function is interpo- rw E½Lðy; yw Þ
lated from the original one, according to the duration of the
1 (14)
synthesized signature. ¼  lim E½rw F ðx; ydirect ; wÞ  rw F ðx; yw ; wÞ;
!0 
The ranges of the random variables RDi , Rt0;i , Rmi , Rsi , Rus;i ,
Rue;i are partially determined based on a previous study [30], with
which carried out visual Turing tests on synthetic characters
and worked out the admissible ranges of parameter variation: yw ¼ arg max F ðx; y^; wÞ; (15)
y^2Y
varying parameters outside these ranges (in percentage terms)
makes a character unrecognizable. Furthermore, Bhattacharya
et al. [30] concluded that distorted samples corresponding to
50 or 75 percent of the admissible ranges can be regarded as
samples from new writers, whereas < 25% distorted samples
maintain the original handwriting style. We borrow the result
for Rt0;i , Rmi and Rsi ; as for RDi , Rus;i , Rue;i , they are empirically
restricted to within a small range as in [22]. On this basis, for
each genuine signature, two groups of signatures, denoted
below as G1 and G2 respectively, are generated with two differ-
ent distortion levels as shown in Table 1. Signatures in G1 have
lower distortion levels and should rank higher according to
the similarity to the template signature, as compared to those
in G2 . In this context, one can regard G1 as augmented genuine
signatures, and G2 as synthetic skilled forgeries. These two dis-
tortion levels are illustrated in Fig. 3, and some values are
determined experimentally (e.g., we divide lm and hm by 1.5),
leaving room for future improvements.
After signature synthesis, we construct the Sig2Vec model
to learn to rank these synthesized signatures, and optimize
the AP of the signature ranking, as described in the next
subsections.

3.1.2 Average Precision Optimization for Ranking

Given one genuine signature and its synthetic samples, we
compute and rank their similarities and incorporate the AP of
the ranking into the loss function for optimization. Because Fig. 3. Visualization of synthesized signatures from three template sig-
AP is non-differential to the network’s outputs, we resort to natures. Inside each box, the first signature is the template signature,
while signatures in the second and third rows are from G1 and G2 , respec-
the General Loss Gradient Theorem proposed by [18], which tively. The signature images are kindly permitted by the writers (u1010,
provides us with the weight update rule for AP optimization. u1013 and u1028 in DeepSignDB) to be used here.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6477

ydirect ¼ arg max F ðx; y^; wÞ  Lðy; y^Þ: (16) As for ydirect , it can be inferred via a dynamic programming
y^2Y algorithm [18]. To prevent overconfidence of the scoring
function, we add a regularization term to the AP loss, and
In Eq. (14), two directions () of optimization can be used. The obtain the following gradient:
positive direction steps away from worse parameters w, while
the negative direction moves toward better ones.
rw LAP ¼  ½rw F ðx; ydirect ; wÞ  rw F ðx; yw ; wÞ
(21)
In the context of this paper, x refers to the concatenation þ rw F ðx; ydirect ; wÞ:
of the two synthetic signature groups for one given genuine
signature x0 , namely x ¼ G1 [ G2 ¼ fx1 ; . . . ; xN g, and N ¼
jG1 j þ jG2 j; y ¼ f. . . ; yij ; . . .g is the collection of all pairwise Besides the AP loss, we also employ the standard cross-
comparisons, where 8i; j 2 f1; . . . ; Ng; yij ¼ 1 if xi is ranked entropy loss LCE for signature classification. Intuitively,
higher than xj , yii ¼ 0, and yij ¼ 1 otherwise. We define these two losses protect the system from skilled forgery and
the scoring function as follows: random forgery attacks, respectively. The overall optimiza-
tion process is given in Algorithm 2. Note that the optimiza-
1 X
F ðx; y; wÞ ¼ yij ð’ðxi ; wÞ  ’ðxj ; wÞÞ; tion algorithm can also apply to real handwritten signatures
jG1 jjG2 j by simply replacing synthetic signatures in G1 and G2 with
x i 2 G1
genuine signatures and skilled forgeries, respectively.
x j 2 G2
(17) 3.2 Sig2Vec: A CNN Structure for Dynamic
where
Signatures
fðx0 ; wÞ  fðxi ; wÞ The Sig2Vec model extracts holistic representations from time
’ðxi ; wÞ ¼ ; (18) functions of dynamic signatures, as shown in Fig. 4. It is
jfðx0 ; wÞjjfðxi ; wÞj
loosely inspired by VGGNet [35] and feature pyramid net-
and fð; wÞ computes signature representations using the works [36]. The backbone network consists of six convolu-
Sig2Vec model with learnable parameters w. The scoring tional layers, each followed by the scaled exponential linear
function F ðx; y; wÞ is inspired by [33] and ’ðxi ; wÞ measures unit (SELU) [37], and three max pooling layers. After the back-
the cosine similarity of representations from samples x0 and bone, two SP modules are employed to obtain the signature
xi . We notice that similar definitions to Eqs. (17) and (18) representation. The first SP module pools the output sequence
have been used for few-shot learning [34]. of the sixth convolutional layer. The second SP module is on
top of a fused feature sequence, which fuses the outputs of the
Algorithm 2. Network Optimization for Learning fourth and sixth convolutional layers
Dynamic Signature Representations
Ffused ¼ SELUðFC256 ðFconv4 Þ þ UpSample 2 ðFconv6 ÞÞ;
Repeat until converged:
1: Draw a genuine signature x0 from a random class (i.e., each (22)
individual is a different class) and synthesize G1 and G2
where Fconv4 2 RL=4 128 , Fconv6 2 RL=8 256 , L is the length of
according to Algorithm 1.
2: Forward pass of network.
input time functions, and 00=400 and 00=800 denote the down-
3: Infer yw and ydirect in Eqs. (15) and (16). sampling rates of the layers. Here, we ignore the batch size
4: Compute AP loss gradient rw LAP as in Eq. (21), and cross- for notational convenience. During the training stage, the
entropy loss gradient rw LCE . AP loss and cross-entropy are applied on top of the signa-
5: Update network (weight decay and momentum are omitted ture representation for network optimization. During the
for simplicity): w w  aðrw LAP þ rw LCE Þ. test stage, the trained network is used to extract signature
representations for the verifier.
Further, let p ¼ rankðyÞ 2 f0; 1gjG1 jþjG2 j be a vector con-
structed by sorting the data points according to the ranking 3.2.1 Selective Pooling
defined by y, such that pj ¼ 1 if the jth data point belongs to The SP module is designed to pool the feature sequence
G1 and pj ¼ 0 otherwise. Then, given ground truth and pre- (e.g., Fconv6 ) of a 1D convolutional layer into a fixed-length
dicted configurations y and y^, the AP loss is vector. It is inspired by the multi-head attention [24], and is
illustrated in the righthand side of Fig. 4. The essential dif-
1 X
Lðy; y^Þ , LAP ðp; p^Þ ¼ 1  Prec@j; ference is that we introduce learnable queries to guide the
jG1 j attention process. Specifically, we first use NSP linear layers
j : p^j ¼ 1 (19)
with DSP hidden units to map an input sequence F into
1 j N NSP subspaces: F ! fFi gi¼1
NSP
, where Fi 2 RjF j DSP and jF j
where p^ ¼ rankð^yÞ and Prec@j is the percentage of samples is the length of sequence F . For the ith subspace, we assign
belonging to G1 that are ranked above position j. To com- a DSP -dimensional learnable query wi (i.e., a trainable row
pute the AP loss gradient, we need to infer yw and ydirect in vector) to be responsible for this subspace, and obtain a
Eqs. (15) and (16). For yw , the solution is simple sub-vector from Fi as follows:
  
þ1; if ’ðxi ; wÞ > ’ðxj ; wÞ; wi FiT
ywij ¼ (20) veci ¼ softmax pffiffiffiffiffiffiffiffiffi Fi : (23)
1; otherwise. DSP
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6478 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

Adding another eight time functions, derived from vx , vy and v

and commonly used in literature [7], [38], further improves the
verification accuracy.

3.3 Signature Verifier

We use a distance-based verifier with the same normaliza-
tion technique as in [8]. Specifically, given two signatures xi
and xj , we compute the euclidean distance of their l2 nor-
malized vectors as follows:

fðxi ; wÞ fðxj ; wÞ 2
dðxi ; xj Þ ¼ k  k; (24)
jfðxi ; wÞj jfðxj ; wÞj 2

where fð; wÞ computes a 1024-dimensional vector as men-

tioned previously. Given n template signatures fxk1 ; . . . ; xkn g
from writer k, we compute the average pair-wise distance
Fig. 4. The architecture of the proposed Sig2Vec model as well as the SP as dk (dk ¼ 1 if n ¼ 1). Then, for a test signature xte claiming
module. Each convolutional layer is followed by the SELU activation to be writer k, we compute the following scores:
function [37]. The feature sequence from the sixth convolutional layer is
up-sampled and added to that of the fourth layer by a feature fusion
layer, following [36]. Two SP modules are applied to the sixth convolu- pffiffiffiffiffi
tional layer and the feature fusion layer, respectively, leading to a 1024- ski ðxtest Þ ¼ dðxki ; xte Þ= dk ; 8i 2 f1; . . . ; ng: (25)
dimensional signature representation. Inside each SP module, multi-
head attention [24] with learnable queries is proposed to extract fixed-
length vectors from feature sequences of arbitrary lengths. From these scores, the average score skave and minimum
score skmin are computed. Given a pre-defined threshold c, if
In Eq. (23), the softmax function returns a jF j-dimensional skave þ skmin < c, then xte is considered a genuine signature
vector, and veci is a DSP -dimensional vector. The final out- from writer k; otherwise it is regarded as a forgery. By vary-
put of the SP module is ½vec1 ; . . . ; vecNSP . To encourage ing the threshold c, we can obtain the equal error rates
diverse features, we initialize the queries orthogonally. (EERs) to assess the system performance. Unless mentioned
Compared with traditional average pooling, the SP has two otherwise, a global threshold for all writers is used.
advantages. First, it can assign different weights to different
locations according to their contexts, as described in Eq. (23). 4 EXPERIMENT
Second, different subspaces capture the properties of different 4.1 Database and Protocol
local structures, because the supervising queries are encour-
We conducted extensive experiments on the recently released,
aged to be different. However, the output dimensionality of an
large-scale database DeepSignDB. DeepSignDB consists of
SP module is NSP DSP , which is generally larger than that of
five subsets, namely MCYT [39], BiosecurID [40], Biosecure
average pooling. In this paper, we apply SP modules on Ffused
DS2 [41], e-BioSign DS1 [42], and e-BioSign DS2, and has 69972
and Fconv6 , and set NSP ¼ 16 and DSP ¼ 32. Therefore, the sig-
handwritten signatures from 1526 writers in total. Notably, the
nature representation is a 1024-dimensional vector.
e-BioSign DS1 and DS2 databases were collected with multiple
devices, using either a stylus or finger as the writing input. For
3.2.2 Input Time Functions more detailed information, we refer the readers to [13]. Along
In this paper, we consider the following 12 time functions as with the database, Tolosana et al. [13] designed a standard
the input for the Sig2Vec model: evaluation protocol so that future novel approaches could per-
form a fair comparison to state-of-the-art approaches. For this

Horizontal and vertical velocity: vx , vy .
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi study, we strictly followed this protocol, which mainly

Velocity magnitude: v ¼ v2x þ v2y . involves the following aspects: (1) Number of templates. For

Path-tangent angle: u ¼ arctanðvy =vx Þ. each writer, one or four genuine signatures are used as tem-

cos ðuÞ, sin ðuÞ and pressure p. plates for the verifier, which is termed as 1vs1 or 4vs1, respec-

First-order derivatives of v and u: v, _ u.
_ tively. (2) Imposter scenarios. Both skilled and random

Log curvature radius: r ¼ log ðv=uÞ. _ forgeries are considered to assess the system performance. (3)

Centripetal acceleration: c ¼ v ffi u.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
_ Writing inputs. Both stylus- and finger-written signatures are

Total acceleration: a ¼ v_ 2 þ c2 . considered.
Each time function is normalized to have zero mean and Currently, the training set of the Biosecure DS2 subset is
unit variance. Among these time functions, vx , vy , v and p are still unavailable due to certain legal issues, but the test set is
found to be the most effective: vx , vy and v are directly related already available. Therefore, we only used four subsets for
to the SL parameter extraction algorithm that accepts velocity training, yet still tested on the full test set. When training
profiles as inputs, whereas p provides another information the Sig2Vec model with the learning-by-synthesis method,
source. Although, regarding the synthetic signatures, pressure we only used the genuine signatures. All signatures were
p is simply interpolated from the original one, it is still benefi- resampled at 100 Hz using cubic interpolation if they were
cial to include the pressure information for Sig2Vec training. not collected at this sampling frequency.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6479

TABLE 2
Signature Verification EERs (%), Using Real Handwritten or Synthetic Signatures for Sig2Vec Training With the AP Loss

Databases Skilled forgery Random forgery

4vs1 1vs1 4vs1 1vs1
Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o
pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups
MCYT 1.78 2.38 2.35 4.17 3.84 4.05 0.51 0.21 0.23 1.80 0.55 0.54
BiosecurID 1.12 1.09 1.28 2.20 2.17 2.53 0.71 0.38 0.33 1.30 0.64 0.64
Biosecure DS2 3.11 2.92 3.99 5.11 4.40 4.90 1.64 0.79 0.99 2.85 1.13 1.20
eBS DS1 w1 1.76 2.62 3.33 3.07 5.06 4.80 0.02 0.08 0.29 0.50 0.46 0.69
eBS DS1 w2 2.27 3.61 3.00 3.93 4.56 4.58 0.27 0.59 0.80 1.17 0.48 0.76
eBS DS1 w3 2.76 4.09 3.52 4.48 6.08 6.31 0.46 0.75 0.52 0.83 1.25 1.56
eBS DS1 w4 2.86 2.91 2.33 4.55 5.30 5.14 0.17 0.03 0.02 0.62 0.57 0.78
eBS DS1 w5 3.64 4.96 4.66 4.59 6.37 7.15 0.76 0.72 1.15 1.01 0.76 1.60
eBS DS2 w2 1.17 1.43 2.27 3.22 3.78 5.10 0.79 0.71 0.81 2.07 1.97 2.38
DeepSignDB 2.18 2.54 2.80 3.88 4.08 4.44 0.79 0.48 0.57 1.66 0.84 0.99

Stylus scenario.

4.2 Implementation Details skilled forgery scenario, it achieves higher EERs in the ran-
Regarding signature synthesis, to accelerate training, syn- dom forgery scenario. (2) Signature synthesis with pen-ups
thetic signatures were first generated offline to create two improves the verification accuracy, which suggests the pen-
data pools P 1 and P 2 for each genuine signature. Then, dur- up signal also conveys the writer information in general.
ing training, G1 and G2 were drawn from P 1 and P 2 respec- Nevertheless, signature synthesis without pen-ups is also
tively. We empirically set jP 1 j ¼ jP 2 j ¼ 10, jG1 j ¼ 5 and very effective, and is even better on some subsets. (3) No
jG2 j ¼ 10. Within each batch, we randomly drew four genu- specific model can excel at all subsets, which implies there
ine signatures from four different writers as templates. are some domain shifts between them (due to the acquisi-
Therefore, the batch size was 4 ð1 þ jG1 j þ jG2 jÞ ¼ 64. tion devices and protocols, etc).
As for network optimization, we used the positive direc- As for the finger scenario, following [14], we used the
tion for the AP loss and set  ¼ 5 in Eq. (21). All models same models trained on the stylus-written signatures to test
were optimized using stochastic gradient descent with Nes- the finger-written signatures without any fine-tuning. As the
terov momentum, with the learning rate, momentum and l2 pressure time function in finger-written signatures is nearly
weight decay set as 0.001, 0.9 and 0.00001 respectively. constant, we normalized its maximal value to be 1.0 instead
Label smoothing [43] was used for the cross-entropy and set of performing the zero-mean unit-variance normalization.
as 0.1. We trained the models for M 200 batches, where M Notably, we also normalized the finger-written signature
is the number of writers in the training set, and the final representations with corresponding training set statistics to
models were evaluated. Each experiment was repeated for reduce the domain shifts. Results are reported in Table 3.
five times (using five different random seeds for network ini- We have the following observations. (1) Synthetic and real
tialization) and the average EERs are reported. signatures are also equally effective for Sig2Vec training in
the finger scenario. But interestingly, in contrast to the sty-
lus scenario, training with real handwritten signatures
4.3 Results achieves lower random forgery EERs. (2) Skilled forgery
Following the organization of this paper, we report the EERs of the finger scenario are much higher than those of
effectiveness of the proposed synthesis algorithm, learning the stylus scenario, which is consistent with previous stud-
algorithm, and the SP module in order. ies [8], [14]. Generally, finger-written signatures on mobile
Synthesis. Real handwritten signatures and synthetic signa- devices appear less natural and have larger intra-writer var-
tures with or without synthesized pen-ups were compared to iations. (3) As finger-written signatures do not record the
determine their effects on Sig2Vec training. As mentioned in pen-up information, signature synthesis with pen-ups does
Section 3.1.2, by replacing synthetic signatures in G1 and G2 not improve the accuracy in the finger scenario.
with genuine signatures and skilled forgeries, respectively, the In general, synthetic signatures provide a very promising
same optimization algorithm was used for real handwritten way to train deep learning-based signature verification sys-
signatures. As for signature synthesis without pen-ups, pen- tems and eliminate the requirement of skilled forgeries. Sig-
downs were synthesized individually and connected with vir- nature synthesis with and without pen-ups are both very
tual pen-ups, as described in [25]. effective, and it is recommended to synthesize pen-ups
Results of the stylus scenario are reported in Table 2. We when they are available and reliably recorded.
have three important observations. (1) The proposed learn- Learning. To prove the effectiveness of the AP loss, we first
ing-by-synthesis method is as effective as training with real compared it with binary cross-entropy (BCE). The BCE is a
handwritten signatures. Although training with real hand- standard loss function for binary classification and very com-
written signatures provides considerably lower EERs in the monly used in deep learning-based signature verification [6],
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6480 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

TABLE 3
Signature Verification EERs (%), Using Real Handwritten or Synthetic Signatures for Sig2Vec Training With the AP Loss

Databases Skilled forgery Random forgery

4vs1 1vs1 4vs1 1vs1
Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o
pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups
eBS DS1 w4 8.43 9.09 9.07 12.11 11.90 12.10 0.57 0.62 0.73 1.47 2.09 1.74
eBS DS1 w5 8.75 8.89 7.90 14.31 14.49 12.38 0.65 0.92 0.72 1.85 2.75 2.48
eBS DS2 w5 4.89 4.29 4.03 8.98 8.41 7.32 0.59 1.24 1.59 0.85 1.71 1.71
eBS DS2 w6 5.33 5.61 5.60 8.69 8.69 9.28 0.52 0.38 0.26 0.58 0.90 0.82
DeepSignDB 6.85 6.97 6.65 11.02 10.87 10.27 0.58 0.79 0.83 1.19 1.86 1.69

Finger scenario.

[7], [10], [14]. In general, a Siamese network is trained with complexity of target signatures, some “skilled” forgeries are
BCE to classify whether a signature pair is from the same indeed not so skilled, and can actually be regarded as random
writer or not. To replace the AP loss with BCE, a pair of signa- forgeries. Therefore, the regularization term in Eq. (21)
ture representations were input to a small Siamese network slightly blurs the class boundaries in the feature space and
with two hidden layers and a sigmoid output neuron as in leads to the increase of random forgery EERs. When training
[44]. Within each batch, there were 4 jG1 j ¼ 20 positive pairs Sig2Vec with synthetic signatures, the AP loss does not affect
and 4 jG2 j ¼ 40 negative pairs, which were labeled as (0.9, the random forgery EERs much, because synthetic signatures
0.1) and (0.1, 0.9) respectively. The loss of positive pairs was in G2 are always “skilled”.
doubled for balance. SP Module. Using the proposed learning-by-synthesis
We also experimented with the triplet loss, which has method, we compared the proposed SP module with the
been reported effective [8] in learning signature representa- commonly used average pooling. Specifically, two SP mod-
tions. The distance metric in Eq. (24) was used, and the mar- ules in the Sig2Vec model, as illustrated in Fig. 4, were
gin of the triplet loss was set as 0.1. Within each batch, the replaced by two average pooling layers, leading to 512-
four template signatures were treated as anchors, while sig- dimensional signature representations. For a fairer compari-
natures in G1 and G2 were treated as positive and negative son, we also compared with SP modules with DSP ¼ 32 and
samples, respectively. For each template signature, there NSP ¼ 8. Results of the stylus scenario are given in Table 5.
were jG1 j jG2 j ¼ 50 triplets in total, and only those with We can see that, the proposed SP module achieves a sub-
non-zero losses were selected for training. stantial improvement over average pooling in both skilled
Again, both real handwritten and synthetic signatures forgery and random forgery scenarios. A larger value of
were compared. Results of the stylus scenario are reported in NSP also improves the performance.
Table 4, where we present results on the entire DeepSignDB We applied the distance discriminant ratio (DDR) [38] to
database for brevity. From Table 4, we can see that the AP loss measure the discriminative power of learned representa-
considerably improves over BCE and the triplet loss in the tions for a further analysis. As mentioned in Section 3.2.1,
skilled forgery scenario. Another observation is that, when there are different sub-vectors, or called heads, in the output
training Sig2Vec with real handwritten signatures, the AP vector of an SP module; there are 16 heads in total in two
loss increases the random forgery EERs. The reason is that, SPð32; 8Þ modules. For the ith head, i ¼ 1; . . . ; 16, we com-
depending on the handwriting skills of forgers and the puted its DDR for user k as follows:

TABLE 4
Comparison of the AP Loss, BCE and the Triplet Loss on Sig2Vec Training

Loss functions Skilled forgery Random forgery

4vs1 1vs1 4vs1 1vs1
Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o Real Syn. w/ Syn. w/o
pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups pen-ups
AP 2.18 2.54 2.80 3.88 4.08 4.44 0.79 0.48 0.57 1.66 0.84 0.99
BCE 2.75 3.09 3.22 4.05 4.54 4.73 0.67 0.53 0.56 1.08 0.84 1.01
"0.57 "0.55 "0.42 "0.17 "0.46 "0.29 #0.12 "0.05 #0.01 #0.58 - "0.02
Triplet 2.96 3.00 3.10 4.45 4.47 4.67 0.63 0.51 0.54 0.94 0.83 0.97
"0.78 "0.46 "0.30 "0.57 "0.39 "0.23 #0.16 "0.03 #0.03 #0.72 #0.01 #0.02

Results without arrows are EERs (%) on the DeepSignDB database, while results with arrows indicate the changes of EERs as compared to the AP loss. Stylus
scenario.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6481

TABLE 5
Comparison of the Selective Pooling Module and Traditional
Average Pooling on the DeepSignDB Database (EER, %)

Pooling method Dim. Skilled Random

forgery forgery
4vs1 1vs1 4vs1 1vs1
Average pooling 512 3.11 5.08 0.76 1.40
SP(32, 8) 512 2.89 4.42 0.67 1.02
SP(32, 16) 1024 2.54 4.08 0.48 0.84

Stylus scenario.

ðmDGk  mDF k Þ2
Rki ¼ i i
; (26)
s 2DGk þ s 2DF k
i i
Fig. 6. Visualization of attention weights of the second SP module. We
simply up-sample the attention weights 4 to align with the signature tra-
where DGki is the set of distances, measured by Eq. (24), jectory, and positions with larger attention weights are marked with darker
among genuine signatures of user k, DFik is the set of distan- dots. We can observe that, the first query detects low-curvature bottom
ces between genuine signatures and skilled forgeries, and m loops, the second query detects high-curvature corners, while the third
query detects low-curvature upper loops. The signature images are kindly
and s denote the average and standard deviation respec- permitted by the writer (u1028 in DeepSignDB) to be used here.
tively. We then obtained the median DDR for the ith head
from all users: Ri ¼ medianfR1i ; R2i ; . . . ; Rki ; . . .g. Intuitively,
a higher Ri indicates that the ith head is more discrimina- of shorter durations, as shown in Fig. 7; template signatures
tive of genuine signatures and skilled forgeries. Similarly, with durations of less than two seconds led to about twice
we obtained the median DDR for the full representation as the misclassification frequency in our experiments. This is
Rfull . As for the average pooling layers, we divided their intuitive, as shorter signatures are generally simpler in com-
outputs into 16 (pseudo-)heads and similarly computed the plexity [45] and easier to imitate. Therefore, designing a
median DDRs. In Fig. 5, we illustrate the obtained results. properly complex signature (from the writers’ perspective)
Although individual heads of the SP(32,8) modules are less is also important to achieve a high security level.
discriminative than those of average pooling, they jointly
lead to a higher discrimination. Therefore, the SP module
enables the network to learn more disentangled and diverse 4.4 Comparison on DeepSignDB
representations. We compare our results with the DTW benchmark and the
To intuitively show why the SP module works, in Fig. 6 state-of-the-art TA-RNN model [14]. Although Tolosana
we visualize the attention weights of three trained queries et al. [13] had given the DTW benchmark results, because
in the second SP module for three different signatures from we use different time functions and a different verifier, we
one writer. It is evident that different queries can attend to report our own DTW results. Results of the stylus scenario
different local structures, although no specific supervising and the finger scenario are presented in Tables 6 and 7,
signal is explicitly designed for this purpose. Properties of respectively; our system ROC curves are shown in Fig. 8.
these different local structures are aggregated into different Under both scenarios, SynSig2Vec achieves state-of-the-art
sub-vectors, which leads to more fine-grained representa- or very competitive results on the DeepSignDB database.
tions and an improved verification accuracy. The DTW method is very good at detecting random forger-
Error Analysis. To understand when a signature was mis- ies, and SynSig2Vec as well as the TA-RNN improves over
classified, we used user-specific thresholds to identify those it mainly in detecting skilled forgeries.
template signatures that had non-zero EERs in the 1vs1
skilled forgery, stylus scenario. We found that misclassifica-
tions occur more frequently when template signatures are

Fig. 5. Median DDRs of the learned representations. Ri denotes the Fig. 7. Template signatures with shorter durations lead to verification
median DDR for the ith head, and Rfull for the full representation. errors more frequently.
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6482 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

TABLE 6
Comparison of DTW, TA-RNN and Our Proposed SynSig2Vec on the DeepSignDB Database (EER, %)

Databases Skilled forgery Random forgery

4vs1 1vs1 4vs1 1vs1
DTW TA-RNN Ours DTW TA-RNN Ours DTW TA-RNN Ours DTW TA-RNN Ours
MCYT 2.91 4.3 2.38 6.18 4.4 3.84 0.58 0.2 0.21 1.20 1.1 0.55
BiosecurID 2.23 1.3 1.09 4.10 1.9 2.17 0.64 0.1 0.38 1.20 0.6 0.64
Biosecur DS2 4.64 3.0 2.92 7.65 4.2 4.40 1.70 1.1 0.79 2.50 1.9 1.13
eBS DS1 w1 6.67 4.3 2.62 9.06 5.4 5.06 2.31 0.1 0.08 3.42 2.5 0.46
eBS DS1 w2 8.10 2.9 3.61 8.93 4.0 4.56 1.99 1.4 0.59 2.19 1.7 0.48
eBS DS1 w3 7.86 4.8 4.09 10.00 5.4 6.08 1.43 0.4 0.75 1.96 1.6 1.25
eBS DS1 w4 6.57 5.2 2.91 9.39 5.8 5.30 0.99 0.9 0.03 2.40 1.4 0.57
eBS DS1 w5 9.05 8.0 4.96 10.81 10.6 6.37 2.07 1.4 0.72 1.76 4.1 0.76
eBS DS2 w2 1.71 2.8 1.43 5.54 3.7 3.78 0.71 0.9 0.71 3.21 2.2 1.97
DeepSignDB 4.53 3.3 2.54 7.06 4.2 4.08 1.23 0.6 0.48 1.98 1.5 0.84

Stylus scenario.

TABLE 7
Comparison of DTW, TA-RNN, and Our Proposed SynSig2Vec on the DeepSignDB Database (EER, %)

Databases Skilled forgery Random forgery

4vs1 1vs1 4vs1 1vs1
DTW TA-RNN Ours DTW TA-RNN Ours DTW TA-RNN Ours DTW TA-RNN Ours
eBS DS1 w4 13.30 16.6 9.09 16.44 18.8 11.90 1.84 0.7 0.62 1.40 1.0 2.09
eBS DS1 w5 13.70 13.3 8.89 17.36 16.4 14.49 1.49 0.7 0.92 1.54 1.7 2.75
eBS DS2 w5 6.69 10.0 4.29 11.49 9.8 8.41 0.44 1.4 1.24 0.80 2.3 1.71
eBS DS2 w6 8.93 5.7 5.61 13.66 8.4 8.69 0.29 1.4 0.38 1.27 1.7 0.90
DeepSignDB 10.66 11.3 6.97 14.74 13.8 10.87 1.02 1.0 0.79 1.25 1.8 1.86

Finger scenario.

It should be noticed that, compared with the TA-RNN number of template signatures. This is mainly because we use
model, we do not have the training set of the Biosecure DS2 a different verifier as described in Section 3.3, especially the
subset, which occupies the largest proportion of the full distance normalization method in Eq. (25). Distance normali-
DeepSignDB database. Therefore, we can safely deduce zation in Eq. (25) is found to play an important role in reduc-
that, if the training set of Biosecure DS2 is made available in ing the global threshold EERs when multiple templates are
future, the performance of SynSig2Vec can be further available, as it adjusts the distances properly according to
improved. More importantly, SynSig2Vec does not require intra-writer statistics.
training with skilled forgeries. This property is significant
when developing new deep learning-based dynamic signa-
ture verification systems, because skilled forgeries are typi- 4.5 Comparisons on MCYT-100 and SVC-Task2
cally very difficult to collect. Because the DeepSignDB database was released very recently
Another interesting observation is that, SynSig2Vec (com- and few results have been reported on this database, we fur-
pared with TA-RNN [14]) and our DTW benchmark (com- ther compare SynSig2Vec with published state-of-the-art
pared with Tolosana’s [13]) benefit more from an increasing methods on two commonly used benchmark databases,
namely MCYT-100 [39] and SVC-Task2 [52], in Table 8. The
MCYT-100 database has been included in the DeepSignDB
evaluation set, and we have reported the results in the above
experiments. The SVC-Task2 database contains 40 writers,
with 20 genuine and 20 forged signatures per writer. Because
SVC-Task2 does not record pen-ups, we connected pen-
downs with virtual pen-ups, and used the same models
trained with signature synthesis without pen-ups to extract
signature representations for the verifier. For each of the 40
writers, one or five genuine signatures were randomly
selected as the templates, while the rest signatures were used
for testing; this procedure was repeated 10 times for each
Fig. 8. System ROC curves on the DeepSignDB database. model (recall that we trained five models using five different
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6483

TABLE 8
Comparison of EERs (%) With Published State-of-the-Art Methods on the MCYT-100 and the SVC-Task2 Databases

Databases Methods Number of templates Skilled forgery Random

forgery
thg thu thg thu
MCYT-100 SRSS based on SL model [22] 1 13.56 - 4.04 -
Stroke-RNN [10] 1 10.46 - - -
SynSig2Vec, ours 1 3.84 1.59 0.55 0.12
Symbolic representation [46] 5 5.70 2.20 2.3 1.0
DTW cost matrix information [47] 5 2.76 1.15 - -
DTW with SCC [48] 5 - 2.15 - -
Recurrent adaptation networks [8] 5 1.81 - 0.24 -
DeepDTW [12] 5 2.40 - - -
Single-template strategy [49] 5 - 1.28 - -
Single-template strategy+LS-DTW [50] 5 - 0.72 - 0.07
Two-tier ensemble [51] 20 - 2.84 - -
SynSig2Vec, ours 4 2.38 0.96 0.21 0.06
SVC-Task2 SRSS based on SL model [22] 1 18.25 - 1.90 -
SynSig2Vec, ours 1 12.16 5.83 0.57 0.14
DTW cost matrix information [47] 5 7.80 2.53 - -
DTW with SCC [48] 5 - 2.63 - -
Single-template strategy [49] 5 - 2.98 - -
Single-template strategy+LS-DTW [50] 5 - 2.08 - 0.11
Two-tier ensemble [51] 15 - 2.2 - -
SynSig2Vec, ours 5 3.88 2.08 0.17 0.02

Global and user-specific thresholds are denoted as thg and thu , respectively.

random seeds), and the average results are reported. Note that 5 CONCLUSION AND FUTURE WORK
we used the first genuine signature from each of the remain-
In this paper, we propose a novel learning-by-synthesis
ing writers as random forgeries.
method and the Sig2Vec model, jointly termed SynSig2Vec for
From Table 8, we can see that SynSig2Vec achieves very
short, to learn dynamic signature representations. Specifically,
promising results on both MCYT-100 and SVC-Task2 data-
we synthesize various signatures from template signatures
bases. It should be noted that SynSig2Vec was trained on
based on the kinematic theory of rapid human movements
the large-scale DeepSignDB database, yet other studies con-
and its SL model, construct the Sig2Vec model to learn to rank
structed their models by using only the target database,
these synthesized signatures, and optimize the average preci-
namely MCYT-100 or SVC-Task2. On the one hand, this
sion of the signature ranking. Through this novel training
demonstrates that SynSig2Vec indeed learns very effective
method, the Sig2Vec model can learn very effective signature
and general signature representations. On the other hand, if
representations for verification. A highlight of SynSig2Vec is
such a large-scale training set is unavailable, the perfor-
that, it requires only genuine signatures for training, which is
mance of SynSig2Vec may decrease. In the conference paper
more practical for developing deep learning-based systems.
[25], we performed a cross validation on the SVC-Task2
Furthermore, to the best of our knowledge, it is the first
database, and the results were somewhat worse due to a
method that uses pure CNNs to extract holistic representations
smaller training set. Because different methods generally
from raw time functions of dynamic signatures. The proposed
use very different protocols, the results are not fully compa-
SynSig2Vec achieves state-of-the-art results on the largest
rable. Therefore, we suggest future researchers to conduct
dynamic signature database to date, DeepSignDB.
experiments on the DeepSignDB database and follow
A limitation of SynSig2Vec is that the signature distortion
exactly the same experiment protocol [13].
levels are determined quite empirically. Although the con-
figurations in Table 1 work well in practice, in future work
4.6 Runtime Speed it is worthy of investigation to explore the effects of signa-
On the MCYT-100 subset, with a batch size of 50, a GTX ture distortion levels on the final verification performance.
1080Ti GPU, an Intel(R) Core(TM) i5-8600K CPU and a Another limitation is that, currently we consider only the
Python3.7 + PyTorch1.6 environment, it takes 1.30 seconds kinematic properties during signature synthesis, and ignore
for Sig2Vec feature extraction plus verification of genuine the signature shape and structure information. In future
signatures and skilled forgeries, whereas it takes 29.51 sec- work, it is possible to incorporate the virtual stroke targets
onds for the DTW method (the fastDTW implementation in into the signature synthesis algorithm as in [23], so that we
https://github.com/slaypni/fastdtw). Owing to its recur- can directly control the signature shape and add more
sive nature, DTW cannot run in parallel, as opposed to Sig2- diversity to the generated signatures.
Vec. The same is true for RNN-based methods, e.g., TA- Currently, we use the same system for signatures of dif-
RNN [14]. Therefore, Sig2Vec is both accurate and efficient. ferent durations and complexities. As suggested in [53], it is
Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
6484 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 44, NO. 10, OCTOBER 2022

possible to first detect the signature complexities, and then [17] R. Plamondon, “A kinematic theory of rapid human movements,”
Biol. Cybern., vol. 72, no. 4, pp. 295–307, 1995.
apply different verification systems for different complexi- [18] Y. Song et al., “Training deep neural networks via direct loss mini-
ties. Such a divide-and-conquer strategy is expected to mization,” in Proc. Int. Conf. Mach. Learn., 2016, pp. 2169–2177.
improve the verification accuracy. [19] J. Galbally, R. Plamondon, J. Fierrez, and J. Ortega-Garcia ,
Notably, based on the SL model, a biometric attack case “Synthetic on-line signature generation. part I: Methodology and
algorithms,” Pattern Recognit., vol. 45, no. 7, pp. 2610–2621, 2012.
is presented in [54], where a genuine-like signature can be [20] M. A. Ferrer, M. Diaz, C. Carmona-Duarte , and A. Morales, “A
reconstructed from a forgery to successfully attack an auto- behavioral handwriting model for static and dynamic signature
matic signature verifier. Such an attack-by-synthesis scheme synthesis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 6,
pp. 1041–1053, Jun. 2017.
poses a new security threat and is worth further research.
[21] M. A. Ferrer et al., “Static and dynamic synthesis of bengali and
Intuitively, if a system is explicitly trained to identify these devanagari signatures,” IEEE Trans. Cybern., vol. 48, no. 10,
synthesized “genuine” samples as forgeries, then it will be pp. 2896–2907, Oct. 2018.
much more robust against these malicious attacks. [22] M. Diaz, A. Fischer, M. A. Ferrer, and R. Plamondon, “Dynamic
signature verification system based on one real signature,” IEEE
Trans. Cybern., vol. 48, no. 1, pp. 228–239, Jan. 2018.
ACKNOWLEDGMENTS [23] M. A. Ferrer, M. Diaz, C. Carmona-Duarte , and R. Plamondon,
“Generating off-line and on-line forgeries from on-line genuine sig-
This work was supported in part by the National Natural Sci- natures,” in Proc. Int. Carnahan Conf. Secur. Technol., 2019, pp. 1–6.
ence Foundation of China (NSFC under Grants 61936003 and [24] A. Vaswani et al., “Attention is all you need,” in Proc. Adv. Neural
61771199), and the Natural Science Foundation of Guang- Inf.Process. Syst., 2017, pp. 5998–6008.
dong Province (GD-NSF under Grant 2017A030312006). [25] S. Lai, L. Jin, L. Lin, Y. Zhu, and H. Mao, “SynSig2Vec: Learning rep-
resentations from synthetic dynamic signatures for real-world ver-
ification,” Proc. AAAI Conf. Artif. Intell., vol. 34, no. 01, pp. 735–742,
REFERENCES 2020.
[1] R. Plamondon and G. Lorette, “Automatic signature verification [26] K. Sundararajan and D. L. Woodard, “Deep learning for bio-
and writer identification—the state of the art,” Pattern Recognit., metrics: A survey,” ACM Comput. Surv., vol. 51, no. 3, pp. 1–
vol. 22, no. 2, pp. 107–131, 1989. 34, 2018.
[2] R. Plamondon and S. N. Srihari,“Online and off-line handwriting [27] S. Chopra, R. Hadsell, and Y. LeCun , “Learning a similarity
recognition: A comprehensive survey,” IEEE Trans. Pattern Anal. metric discriminatively, with application to face verification,”
Mach. Intell., vol. 22, no. 1, pp. 63–84, Jan. 2000. in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2005, vol. 1,
[3] D. Impedovo and G. Pirlo, “Automatic signature verification: The pp. 539–546.
state of the art,” IEEE Trans. Syst., Man, Cybern., Appl. Rev., vol. 38, [28] C. Sekhar, P. Mukherjee, D. S. Guru, and V. Pulabaigari, “OSVNet:
no. 5, pp. 609–635, Sep. 2008. Convolutional Siamese network for writer independent online
[4] M. Diaz, M. A. Ferrer, D. Impedovo, M. I. Malik, G. Pirlo, and signature verification,” in Proc. Int. Conf. Document Anal. Recognit.,
R. Plamondon, “A perspective analysis of handwritten signature 2019, pp. 1470–1475.
technology,” ACM Comput. Surv., vol. 51, no. 6, 2019, Art. no. 117. [29] S. Hochreiter and J. Schmidhuber, “Long short-term memory,”
[5] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
vol. 521, no. 7553, pp. 436–444, 2015. [30] U. Bhattacharya, R. Plamondon, S. D. Chowdhury, P. Goyal, and
[6] K. Ahrabian and B. Babaali, “Usage of autoencoders and Siamese S. K. Parui, “A sigma-lognormal model-based approach to gener-
networks for online handwritten signature verification,” Neural ating large synthetic online handwriting sample databases,” Int. J.
Comput. Appl., vol. 31, pp. 9321–9334, 2019. Document Anal. Recognit., vol. 20, no. 3, pp. 155–171, 2017.
[7] R. Tolosana, R. Vera-Rodriguez , J. Fierrez, and J. Ortega-Garcia , [31] C. O’Reilly and R. Plamondon, “Development of a Sigma–Lognor-
“Exploring recurrent neural networks for on-line handwritten sig- mal representation for on-line signatures,” Pattern Recognit.,
nature biometrics,” IEEE Access, vol. 6, pp. 5128–5138, 2018. vol. 42, no. 12, pp. 3324–3337, 2009.
[8] S. Lai and L. Jin, “Recurrent adaptation networks for online signa- [32] M. A. Ferrer, M. Diaz, C. Carmona-Duarte , and R. Plamondon,
ture verification,” IEEE Trans. Inf. Forensics Secur., vol. 14, no. 6, “iDeLog: Iterative dual spatial and kinematic extraction of sigma-
pp. 1624–1637, Jun. 2019. lognormal parameters,” IEEE Trans. Pattern Anal. Mach. Intell.,
[9] C.-Y. Park, H.-G. Kim, and H.-J. Choi, “Robust online signature vol. 42, no. 1, pp. 114–125, Jan. 2020.
verification using long-term recurrent convolutional network,” in [33] Y. Yue, T. Finley, F. Radlinski, and T. Joachims, “A support vector
Proc. IEEE Int. Conf. Consumer Electron., 2019, pp. 1–6. method for optimizing average precision,” in Proc. 30th Annu. Int.
[10] C. Li, X. Zhang, F. Lin, Z. Wang, L. Jun’E, R. Zhang, and H. Wang, “A ACM SIGIR Conf. Res. Develop. Inf. Retrieval, 2007, pp. 271–278.
stroke-based RNN for writer-independent online signature ver- [34] E. Triantafillou, R. Zemel, and R. Urtasun, “Few-shot learning
ification,” in Proc. Int. Conf. Document Anal. Recognit., 2019, pp. 526–532. through an information retrieval lens,” in Proc. Adv. Neural Inf.
[11] X. Wu, A. Kimura, S. Uchida, and K. Kashino, “Prewarping Sia- Process. Syst., 2017, pp. 2255–2265.
mese network: Learning local representations for online signature [35] K. Simonyan and A. Zisserman, “Very deep convolutional net-
verification,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process., works for large-scale image recognition,” in Proc. Int. Conf. Learn.
pp. 2467–2471. Representations, May 2015.
[36] T. Lin, P. DollAar, e R. Girshick, K. He, B. Hariharan, and
[12] X. Wu, A. Kimura, B. K. Iwana, S. Uchida, and K. Kashino, “Deep
dynamic time warping: End-to-end local representation learning S. Belongie, “Feature pyramid networks for object detection,” in
for online signature verification,” in Proc. Int. Conf. Document Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 936–944.
Anal. Recognit., 2019, pp. 1103–1110. [37] G. Klambauer, T. Unterthiner, A. Mayr, and S. Hochreiter, “Self-
[13] R. Tolosana, R. Vera-Rodriguez , J. Fierrez, A. Morales, and normalizing neural networks,” in Proc. Adv. Neural Inf. Process.
J. Ortega-Garcia , “Do you need more data? The DeepSignDB on- Syst., 2017, pp. 971–980.
line handwritten signature biometric database,” in Proc. Int. Conf. [38] M. Martinez-Diaz , J. Fierrez, R. P. Krish, and J. Galbally, “Mobile
Document Anal. Recognit., 2019, pp. 1143–1148. signature verification: Feature robustness and performance
[14] R. Tolosana, R. Vera-Rodriguez , J. Fierrez, and J. Ortega-Garcia , comparison,” IET Biometrics, vol. 3, no. 4, pp. 267–277, 2014.
“Deepsign: Deep on-line signature verification,” IEEE Trans. Bio- [39] J. Ortega-Garcia et al., “MCYT baseline corpus: A bimodal biomet-
metrics, Behav., Identity Sci., vol. 3, no. 2, pp. 229–239, Apr. 2021. ric database,” IEEE Proc. Vis., Image Signal Process., vol. 150, no. 6,
[15] D. Guru and H. Prakash, “Online signature verification and recogni- pp. 395–401, Dec. 2003.
tion: An approach based on symbolic representation,” IEEE Trans. [40] J. Fierrez et al., “BiosecurID: A multimodal biometric database,”
Pattern Anal. Mach. Intell., vol. 31, no. 6, pp. 1059–1073, Jun. 2009. Pattern Anal. Appl., vol. 13, no. 2, pp. 235–246, 2010.
[16] N. Sae-Bae and N. Memon, “Online signature verification on [41] J. Ortega-Garcia et al., “The multiscenario multienvironment bio-
mobile devices,” IEEE Trans. Inf. Forensics Secur., vol. 9, no. 6, secure multimodal database (BMDB),” IEEE Trans. Pattern Anal.
pp. 933–947, Jun. 2014. Mach. Intell., vol. 32, no. 6, pp. 1097–1111, Jun. 2010.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.
LAI ET AL.: SYNSIG2VEC: FORGERY-FREE LEARNING OF DYNAMIC SIGNATURE REPRESENTATIONS BY SIGMA LOGNORMAL-BASED... 6485

[42] R. Tolosana, R. Vera-Rodriguez , J. Fierrez, A. Morales, and Lianwen Jin received the BS degree from the Uni-
J. Ortega-Garcia , “Benchmarking desktop and mobile handwrit- versity of Science and Technology of China, Anhui,
ing across COTS devices: The e-BioSign biometric database,” PloS China, in 1991 and the PhD degree from the South
One, vol. 12, no. 5, 2017, Art. no. e0176792. China University of Technology, Guangzhou, China,
[43] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, in 1996 . He is currently a professor at the College of
“Rethinking the Inception architecture for computer vision,” in Proc. Electronic and Information Engineering, South
IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 2818–2826. China University of Technology. He has authored
[44] S. Lai and L. Jin, “Learning discriminative feature hierarchies for more than 100 scientific papers. His research inter-
off-line signature verification,” in Proc. 16th Int. Conf. Front. Hand- ests include handwriting analysis and recognition,
writing Recognit., 2018, pp. 175–180. image processing, machine learning, and intelligent
[45] N. Sae-Bae , N. Memon, and P. Sooraksa, “Distinctiveness, com- systems. He was the recipient of the New Century
plexity, and repeatability of online signature templates,” Pattern Excellent Talent Program of MOE Award and the Guangdong Pearl River
Recognit., vol. 84, pp. 332–344, 2018. Distinguished Professor Award. He is currently a member of the IEEE
[46] D. Guru, K. Manjunatha, S. Manjunath, and M. Somashekara, Computational Intelligence Society, IEEE Signal Processing Society, and
“Interval valued symbolic representation of writer dependent fea- IEEE Computer Society.
tures for online signature verification,” Expert Syst. Appl., vol. 80,
pp. 232–243, 2017.
[47] A. Sharma and S. Sundaram, “On the exploration of information Yecheng Zhu received the BS degree from the
from the DTW cost matrix for online signature verification,” IEEE South China University of Technology, in 2018.
Trans. Cybern., vol. 48, no. 2, pp. 611–624, Feb. 2018. He is currently working toward the master’’s
[48] X. Xia, Z. Chen, F. Luan, and X. Song, “Signature alignment based degree in signal and information processing. His
on GMM for on-line signature verification,” Pattern Recognit., research interests include machine learning,
vol. 65, pp. 188–196, 2017. handwriting analysis, and computer vision.
[49] M. Okawa, “Online signature verification using single-template
matching with time-series averaging and gradient boosting,” Pat-
tern Recognit., vol. 102, 2020, Art. no. 107227.
[50] M. Okawa, “Time-series averaging and local stability-weighted
dynamic time warping for online signature verification,” Pattern
Recognit., vol. 112, 2021, Art. no. 107699. Zhe Li received the BS degree in electronics and
[51] P. Bhowal, D. Banerjee, S. Malakar, and R. Sarkar, “A two-tier information engineering, in 2018, from the South
ensemble approach for writer dependent online signature ver- China University of Technology, where he is cur-
ification,” J. Ambient Intell. Humanized Comput., pp. 1–20, 2021. rently working toward the PhD degree in informa-
[52] D.-Y. Yeung et al., “SVC2004: First international signature verifica- tion and communication engineering. His research
tion competition,” in Proc. Int. Conf. Biometric Authentication, 2004, interests include deep learning, handwriting analy-
pp. 16–22. sis and recognition, and sequence modeling.
[53] R. Vera-Rodriguez et al., “DeepSignCX: Signature complexity detec-
tion using recurrent neural networks,” in Proc. Int. Conf. Document
Anal. Recognit., 2019, pp. 1326–1331.
[54] M. A. Ferrer, M. Diaz, C. Carmona-Duarte , and R. Plamondon, “A
biometric attack case based on signature synthesis,” in Proc. Int.
Luojun Lin received the BS degree in electronic
Carnahan Conf. Secur. Technol., 2018, pp. 1–6.
and information engineering from Yunnan Univer-
sity, Kunming, China in 2014, and the PhD degree
Songxuan Lai received the BS degree in elec-
in information and communication engineering
tronics and information engineering, in 2016, from the South China University of Technology,
from the South China University of Technology, Guangzhou, China, in 2020. She is currently at the
where he is currently working toward the PhD School of Mathematics and Computer Science,
degree in information and communication engi- Fuzhou University. Her research interests include
neering. He has authored several papers in high-
image analysis, pattern recognition, and machine
impact conferences and journals, including Inter-
learning.
national Conference on Document Analysis and
Recognition, Association for the Advancement of
Artificial Intelligence, and the IEEE Transactions " For more information on this or any other computing topic,
on Information Forensics and Security. His
research interests include machine learning, OCR systems, handwriting please visit our Digital Library at www.computer.org/csdl.
analysis and recognition, and signature verification.

Authorized licensed use limited to: VTU Consortium. Downloaded on October 03,2023 at 16:24:42 UTC from IEEE Xplore. Restrictions apply.