MorDeephy: Face Morphing Detection via Fused Classification
Iurii Medvedev
1 a
, Farhad Shadmand
1 b
and Nuno Gonc¸alves
1,2 c
1
Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal
2
Portuguese Mint and Official Printing Office (INCM), Lisbon, Portugal
Keywords:
Face Morphing Detection, Face Recognition, Deep Learning, Convolutional Neural Networks, Classification.
Abstract:
Face morphing attack detection (MAD) is one of the most challenging tasks in the field of face recognition
nowadays. In this work, we introduce a novel deep learning strategy for a single image face morphing de-
tection, which implies the discrimination of morphed face images along with a sophisticated face recognition
task in a complex classification scheme. It is directed onto learning the deep facial features, which carry in-
formation about the authenticity of these features. Our work also introduces several additional contributions:
the public and easy-to-use face morphing detection benchmark and the results of our wild datasets filtering
strategy. Our method, which we call MorDeephy, achieved the state of the art performance and demonstrated
a prominent ability for generalizing the task of morphing detection to unseen scenarios.
1 INTRODUCTION
The development of deep learning techniques in the
last decades allowed to achieve evident advances in
the area of face recognition. However, evolved and
sophisticated techniques for performing the presenta-
tion attacks continue to appear, which require the de-
velopment of new protection solutions.
Face morphing is one such image manipulating
technique. It is usually performed by blending sev-
eral (usually two) digital face images and allows to
match different persons with this synthetic image that
contains characteristics from both faces. While being
simple to implement, face morphing poses the secu-
rity risks of issuing an identification document that
may be validated for two or more persons. Presenta-
tion attacks with face morphing usually can be hardly
detected by humans, which usually perform poorly in
matching unfamiliar faces on photos of ID and travel
documents (Medvedev et al., 2020) and by face recog-
nition software in ABC (automatic border control)
systems (Ferrara et al., 2014).
In the last years, face morphing has become a mat-
ter of research interest in academia (NIST, 2020) and
industry (Research and Service, 2020). Morphing
detection methods in facial biometric systems may
be distinguished into two pipelines depending on the
processing scenario. In no-reference morphing attack
a
https://orcid.org/0000-0003-2372-9681
b
https://orcid.org/0000-0003-4399-4845
c
https://orcid.org/0000-0002-1854-049X
detection algorithm receives a single image, where
morphing is detected. In practice, these methods are
directed to mitigate risks related to the false accep-
tance of manipulated images in the enrollment pro-
cess. The authentic document, which is generated
with a successfully accepted forged image, may fur-
ther help to deceive the face recognition system.
The differential morphing detection implies addi-
tional live data acquisition from an authentication sys-
tem which gives the reference information for the de-
tection algorithm. This scenario usually takes place
while passing an Automated Border Control (ABC)
system, when the recently enrolled image (which is
already accepted and printed on the ID Document) is
tested against morphing detection.
First morphing detection solutions relied on the
behaviour of local image characteristics (like tex-
ture, noise). Recent approaches usually employ deep
learning computer vision tools. However, many of
these methods utilize a straightforward learning strat-
egy that is limited by binary classification or contrast
learning, which in our opinion is not optimal for a task
of face morphing detection and may lead to various
convergence problems.
In this work, we introduce a novel deep learn-
ing method for single image face morphing detec-
tion, which incorporates sophisticated face recogni-
tion tasks and implies utilizing a combined classifi-
cation scheme (discussed in Section 3). We propose
the new strategy of face morphs classification, which
is used for the purpose of face morphing detection.
Medvedev, I., Shadmand, F. and Gonçalves, N.
MorDeephy: Face Morphing Detection via Fused Classification.
DOI: 10.5220/0011606100003411
In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 193-204
ISBN: 978-989-758-626-2; ISSN: 2184-4313
Copyright
c
2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
193
This novel approach leads to the usage of CNN cou-
pling, which is accompanied by a sophisticated label-
ing strategy and a specific data mining technique.
Also, we develop the public face morphing detec-
tion benchmark, which is designed to be adaptive to
the developer needs and at the same time to be simple
for comparison of algorithms of different developers.
As an additional contribution, we introduce the results
of our datasets filtering strategy (image name lists),
which is described in Section 4.1.
Regarding the limitations of the work, it is im-
portant to note that at the current stage we focus
on single image morphing detection. Also, we do
not take into account redigitalized face images (by
printing/scanning), however we demonstrate that our
method allow to generalize the detection onto this
case. At the same time, we are limited to uti-
lizing landmark-based methods for performing face
morphing. GAN (Generative adversarial Network)
based methods require large computational resources
(namely for projecting images to latent space) and
at the same time, face recognition systems are less
vulnerable to presentation attacks with GAN morphs,
rather than to landmark-based morphs (Venkatesh
et al., 2020). However, we intend to cover those limi-
tations in further research.
2 RELATED WORK
To introduce our methodology, we need to discuss re-
cent advances in face morphing, face morphing detec-
tion (focusing on the no-reference scenario) and face
recognition.
2.1 Face Morphing
The generic pipeline of creating face morph from
original images includes the following steps: face fea-
tures extraction features averaging generating
morphed image from averaged features optional
restoring image context (namely background).
Landmark based approaches, first introduced by
Ferrara et al. (Ferrara et al., 2014), follow this
pipeline straightforwardly in the image spatial do-
main by the face landmark alignment, image warp-
ing and blending. Different reported morphing algo-
rithms employ variations of this strategy (Laboratory,
2018; LLC, 2010; Satya, 2016).
With recent advances in generative deep learning
approaches, several face morphing methods, which
utilize deep latent feature domain, were proposed.
The above face morphing pipeline may adapt var-
ious deep learning tools like variational autoencoders
(VAE) (Damer et al., 2018) or generative adversar-
ial networks (GANs) (Venkatesh et al., 2020; Zhang
et al., 2020).
2.2 Face Morphing Detection
Single image (no-reference) face morphing detection
algorithms usually utilise local image information and
image statistics.
Various morphing detection approaches em-
ploy Binarized Statistical Image Features (BSIF)
(Raghavendra et al., 2016), Photo Response Non-
Uniformity (PRNU), known as sensor noise (De-
biasi et al., 2018; Scherhag et al., 2019), texture
features (Ramachandra et al., 2019), local features
in frequency and spatial image domain (Neubert
et al., 2019) or complex combination of these features
(Lorenz et al., 2021; Scherhag et al., 2017).
Several deep learning approaches for no-reference
case were proposed. For face morphing detection
these approaches usually follow binary classification
of pretrained face recognition features (Raghavendra
et al., 2017), which may be finetuned (Seibold et al.,
2017; Ferrara et al., 2021) or utilized in a combina-
tion with local texture characteristics (Wandzik et al.,
2018). Damer et al. (Damer et al., 2021) introduced a
better regularized strategy for morphing detection by
replacing the trivial binary classification with pixel-
wise supervision. Aghdaie et al. (Aghdaie et al.,
2021) adopted the attention mechanism which is con-
trolled by wavelet decomposition.
Differential face morphing detection is a less chal-
lenging task and security risks in this scenario indeed
may be combated by increasing the discriminabil-
ity of face deep representation, which is utilized for
recognition.
Several approaches for differential detection was
recently proposed. Scherhag et al. (Scherhag
et al., 2020) followed the classification of pretrained
deep features in differential scenario. Borghi et al.
(Borghi et al., 2021b) performed differential morph-
ing detection by finetuning the pretrained networks
in a complex setup with identity verification and ar-
tifacts detection blocks. Rather different approach to
the differential scenario was introduced by Ferrara et
al. (Ferrara et al., 2018) who proposed an approach
to revert morphing by retouching the testing face im-
age with a trusted live capture to reveal the identity of
the legitimate document owner.
In comparison to the considered approaches, we
propose to focus our method on learning the authen-
ticity of deep face features, regularizing the morphing
detection with a delicate face recognition task.
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
194
Categorical
Batch 1
Morphing
y = n000002
.
y = n005206
..
(bona
de)
Feature
Layer f
.
.
.
Feature
Layer f
.
.
.
Dot Product
& sigmoid
activation
Cross entropy
.
.
.
.
.
.
Batch 2
(face morph)
y = n000002
y = n005206
..
.
Class
Layer 1
Class
Layer 2
Categorical
Cross entropy
y = n000002
..
y = n000002
.
Binary
Cross entropy
...
De ne
cross label t
Weighted
Sum
L
.
..
W,
.
W,
..
b
.
b
..
First Network
Second Network
Figure 1: Schematic of the proposed method. For simplicity of visualization batch contains a single image. Labels ˙y and ¨y are
indicated by names, when the real setup utilizes their numerical index value, which is encoded to one-hot vector.
2.3 Face Recognition
Modern face recognition approaches rely on deep
learning tools, which give the ability to learn
highly discriminative features themselves from un-
constrained images. Among several techniques to
perform the tasks of extracting features, the convo-
lutional neural network (CNN) is one of the most
efficient for the pattern recognition problems (Rus-
sakovsky et al., 2015).
There are several strategies for approaching face
recognition via deep learning. However, all of them
are focused on extracting low-dimensional face repre-
sentation (deep face features) and increasing the dis-
criminative power of that representation.
Metric learning methods are directed on optimiz-
ing the face representation itself through the con-
trast of match/non-match pairs (Chopra et al., 2005;
Schroff et al., 2015). However, for reliable con-
vergence, these methods require enormously large
datasets and sophisticated sample mining techniques.
Another concept (which we indeed follow in our
work) is learning face representation implicitly via
a closed-set identity classification task. Deep net-
works in these methods encapsulate face representa-
tion in the last hidden layer and usually adopt softmax
loss and its modifications for classification (Sun et al.,
2014; Sun et al., 2014; Sun et al., 2015).
Improvement of the performance in this approach
was achieved by various techniques for increasing
intra-class compactness and maximizing inter-class
discrepancy. For example, by applying additional
regularization for pushing intra-class features to their
centre (Wen et al., 2016), or by introducing several
kinds of marginal restrictions for inter-class variance
(Liu et al., 2017; Wang et al., 2018; Deng et al., 2019;
Sun et al., 2020).
Several recent works were directed onto investi-
gating sample specific learning strategies, which are
controlled by sample quality (Tremoc¸o et al., 2021),
hardness (Zeng et al., 2020; Huang et al., 2020), data
augmentation (Shi et al., 2020) or even by treating fa-
cial representation in distributional manner (by speci-
fying sample uncertainty) (Shi and Jain, 2019).
In our work, we consider face morphing detec-
tion from the perspective of face recognition. In the
case of following the approach via identity classifi-
cation, face morphing introduces a problem, since a
face morph image indeed belongs to several identities,
which leads to the ambiguity of proper class labeling.
In this work, we address this issue (Section 4.2) in
search of the method for single image morphing de-
tection.
3 METHODOLOGY
In this section, we describe our technique for single
image morphing detection via deep learning.
In our research, we intended to design a setup,
which will allow to learn high-level deep features, that
also carry information about their authenticity. We
defined two requirements to our setup. First, each in-
MorDeephy: Face Morphing Detection via Fused Classification
195
put image must be classified explicitly and unambigu-
ously. Second, the face morphing detection decision
is made by the behavior of deep face features.
This resulted in the schematic that includes two
backbone CNN based networks that are trained in a
similar manner but biased in a way to discriminate
morphed and bona fide images. Namely, our idea im-
plies training two parallel networks, which consider
bona fide samples similarly and morphed samples dif-
ferently (see Fig. 1). We point out that these networks
do not share the same weights and we do not intend
to train them in a contrastive manner (by matching
positive and negative pairs).
In our setup both networks learn high-level fea-
tures via classification tasks, which are different in
terms of identity labeling of face morphs. First Net-
work labels them by the original identity from the first
source image, the Second Network - by the second
original label.
The extracted features are also explicitly com-
pared by similarity metric (which is the dot product
due to the softmax properties) and the result is clas-
sified according to the ground truth authenticity label
of the image (bona fide/morph). Indeed the variance
of the behavior of the networks is used to make the
detection decision.
The identity classification parts of the training
scheme act as a regularization that retains the facial
discriminability of feature layers. That is why for
identity classification we utilize a standard softmax,
which allows easier convergence in comparison with
its modifications (like ArcFace(Deng et al., 2019)).
Following the common formulation of softmax,
our training process is regularized by the losses:
L
1
=
1
N
N
i
log(
e
˙
W
T
˙y
i
˙
f
i
+
˙
b
˙y
i
C
j
e
˙
f
˙y
j
) (1)
L
2
=
1
N
N
i
log(
e
¨
W
T
¨y
i
¨
f
i
+
¨
b
¨y
i
C
j
e
¨
f
¨y
j
), (2)
where
˙
f
i
,
¨
f
i
denote the deep features of the i th
sample,
{
˙y
i
, ¨y
i
}
are the numerical class indexes of the
i th sample,
˙
W ,
¨
W
and
˙
b,
¨
b
are weights and bi-
ases of last fully connected layer (respectively for the
{
First,Second
}
networks). N is the number of sam-
ples in a batch and C is the total number of classes.
The target driver of the training process explic-
itly discriminates morph/non-morph images. The dot
product of backbones outputs indicates the morphing
detection score. It is activated by sigmoid function,
and the corresponding loss is defined as binary cross-
entropy:
L
3
=
1
N
N
i
t log
1
1+e
D
+ (1 t)log
1
1
1+e
D
,
(3)
where closs-label t is computed by a comparison of
input class labels:
t = abs(sgn( ˙y
i
¨y
i
)), (4)
and D is a dot product of high level features extracted
by First and Second backbones:
D =
˙
f ·
¨
f (5)
The result loss for optimization is combined as a
weighted sum:
L = α
1
L
1
+ α
2
L
2
+ βL
3
(6)
By minimizing this loss in the fused classification
setup, we learn the discriminative facial features that
are explicitly regularized for morphing detection. The
optimal values of weighting coefficients α
1
, α
2
and β
will be obtained empirically in the Section 6.1.
At the testing stage, the identity classification
parts of the network are redundant and may be re-
moved from the setup. The morphing detection de-
cision is made by thresholding the scalar product of
the backbones outputs.
4 DATASETS
The proposed methodology requires the large labeled
face dataset with an accompaniment of morphed im-
ages of identities from this dataset.
The academic community still doesn’t have pub-
lic ID document compliant datasets which are large
enough for efficient training of modern deep networks
(as an example, one of the largest FRGC V2(Phillips
et al., 2005) contains only 50k images and 500
identities). That is why our strategy for this work is
to utilize the wild dataset which is filtered by the cri-
teria of suitability for face morphing. Conceptually
this approach indeed is not novel and was recently uti-
lized in face morphing research (Damer et al., 2019;
Damer et al., 2021). In this work, we introduce a tech-
nique for semi-automatic wild dataset filtering for our
method.
As a source wild dataset we use VGGFace2(Cao
et al., 2018)(3M images, 9k classes, 360 sam-
ples per class, License - CC BY-SA 4.0) due to large
average number of samples per identity (in compari-
son to other popular wild face datasets like CASIA-
WebFace(Yi et al., 2014), MS-Celeb-1M(Guo et al.,
2016; Jin et al., 2018), Glint360K(An et al., 2021),
WebFace260M(Zhu et al., 2021)) in order to have
enough samples per identity after filtering.
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
196
Figure 2: FAR and FRR for result manual quality labeling
of random samples from VGGFace2.
4.1 Wild Dataset Filtering
Our dataset filtering strategy is based on a thresh-
olding by quality metrics. Following Tremoc¸o et
al. (Tremoc¸o et al., 2021) we used a set of qual-
ity scores for labeling the face images in the dataset:
Blur (Bansal et al., 2016), FaceQNet (Hernandez-
Ortega et al., 2019), BRISQUE (Mittal et al., 2012),
Face Illumination (Zhang et al., 2017) and Pose
(Ruiz et al., 2018). This set of scores allow to dis-
criminate and select samples by their natural quality
(Blur, BRISQUE), ID documents suitability (Face-
QNet), face image acquisition parameters (Illumina-
tion, Pose).
Next, we randomly select samples and manually
label them with a binary value (accept/reject). This
acceptance is defined by the criteria of suitability
for application in face morphing (namely by user’s
choice). In our setup, we assure that samples are se-
lected distributively across the quality scores values.
Namely, we split the total quality score range into a
set of sub-ranges and define the minimum number of
samples to be selected from each sub-range. By pro-
ceeding around 4k images in our setup, we harvest
the dependency of FAR (False Acceptance Rate) and
FRR (False Rejection Rate) from quality scores val-
ues (see Fig. 2).
The dataset filtering is then performed with joint
thresholding by those quality metrics. For each score,
we select the threshold at a point of EER (Equal Error
Figure 3: Example of VGGFace2 filtering result. Accepted
images (green box) and Rejected images (red box).
Rate). To combine a filtered dataset we collect sam-
ples, which have a quality score value higher then the
corresponding threshold for each quality metric. As a
result we get the VGGFace2-selected dataset with the
same identity list as original source and around 500k
images (see Fig. 3).
4.2 Morph Dataset Harvesting Strategy
For application in our method, the filtered wild dataset
is needed to be accompanied by a large collection of
face morphs. We automatically generate these im-
ages with our customized landmark-based morphing
approach (with blending coefficient 0.5) (see Fig. 4).
A key requirement for effective learning is to pro-
vide unambiguity of proper class labeling in our train-
ing method. Namely after generating face morph
from two arbitrary samples of the original dataset, the
resulting image indeed belongs to both source iden-
tities. That is why fully random image pairing (for
generating morphs) will result in classification confu-
sion.
To avoid that we utilize the following strategy.
First, we separate the total list of identities into two
disjoint parts, which are attributed to the First and
the Second networks respectively. Next, to generate a
face morph, we randomly pair images from identities
of these list halves. Each generated image is then la-
beled according to the attributed sublist for classifica-
tion by the First and the Second networks. Let us note
that this labeling is made differently for each morphed
image and similarly for bona fide images in both net-
works (see Fig. 1). That is why this technique, which
primarily acts as a regularization, also amplifies the
morphing detection performance.
The separation of a dataset to two disjoint iden-
tity sets is performed to assure that morphed combina-
tions of a particular identity are classified similarly by
the networks. For example, if the dataset include four
identities [A,B,C,D], we split it to [A,B] (assigned for
the First network) and [C,D] (assigned for the Sec-
ond network) and generate possible paired morphs:
MorDeephy: Face Morphing Detection via Fused Classification
197
[AC], [BC], [AD], [BD], which may be unambiguously
classified by both networks by their assigned identity
lists. For instance, in this case, the combination [AB]
is hardly classified. Indeed we assign half of the iden-
tities as main to both networks. Without such separa-
tion, the overall classification loses its regularisation
sense.
Following the above procedure, we generate
VGGFace2-selected-morph dataset, which contains
around 1M morphed images.
+
+
Figure 4: Examples of generated morphs with landmark
based approach. Background is restored by one of the
source images (chosen randomly).
Figure 5: Examples of generated selfmorphs. Images con-
tain blending artefacts but the identity is perceptually re-
tained.
4.3 Selfmorphing
The fully automatic landmark morphing methods may
introduce a number of visible artefacts to the gener-
ated images (like blending artefacts). That is why
without additional regularisation our method will be
biased to learning those artefacts, which is not a real-
istic scenario. Real fraudulent morphs are retouched
with the intention to remove any perceptual artefacts.
To address this problem, we utilize selfmorphs,
which are generated by applying face morphing to im-
ages of the same identity. This concept was indeed
recently introduced by Borghi et al. (Borghi et al.,
2021a) and was used for generating images with vis-
ible artefacts. Then for removing these artefacts the
authors trained the Conditional GAN using original
images as a ground truth reference.
In this work, we utilize selfmorph images to fo-
cus morphing detection onto the deep face features
behaviour, rather than to detecting artefacts. We as-
sume that deep discriminative face features remain af-
ter performing selfmorphing. In the proposed method
schematic (Fig. 1) we consider selfmorphs as bona
fide samples.
We perform a random pairing of samples within
each identity from the VGGFace2-selected and gen-
erate VGGFace2-selected-selfmorph dataset, which
contains around 500k images (see Fig. 5).
5 BENCHMARKING
There are few public benchmarks for evaluating the
performance of morphing detection or morphing re-
sistant algorithms: the NIST FRVT MORPH (NIST,
2022) and FVC-onGoing MAD(Raja et al., 2020;
Dorizzi et al., 2009). Both of these benchmarks
accept no-reference and differential morphing algo-
rithms, however, they are proprietary and are executed
on the maintainer side. Thus they have a number of
submission restrictions.
The straightforward metric for evaluating single
image morphing detection is the dependency of Bona
fide Presentation Classification Error Rate (BPCER)
from Attack Presentation Classification Error Rate
(APCER) (according to ISO/IEC 30107-3 (Inter-
national Organization for Standardization, 2017)),
which may be plotted as a Detection Error Trade-off
(DET) curve.
5.1 Face Morphing Detection
Benchmark
For this work, we intend to develop and provide an
easy-to-use benchmark, which is to be executed on
the developer side (It will be made public in case of
paper acceptance). The existing public benchmarks
provide useful data but usually specify the protocols
for the certain software frameworks (Sarkar et al.,
2020).
Our benchmark intends to provide the function-
ality for estimating the morphing detection perfor-
mance, for generating custom protocols and also for
further comparison of the results from different de-
velopers with existing protocols. At this stage of our
work, we focus on the single image morphing de-
tection with only the usage of public data (however,
we assume the possibility of further adapting private
datasets).
We generate several protocols for single image
morphing detection. Our benchmark is based on
the FRGC-Morphs, FRLL-Morphs (Sarkar et al.,
2020), AMSL (Neubert et al., 2018) and Dustone
datasets(Dustone, 2017). Using this data we com-
bine several benchmark protocols with various types
of face morphs:
protocol-real (3k morphs(Dustone+AMSL)),
which includes morphs with low level of visible
blending artifacts, and imitates realistic presenta-
tion attacks.
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
198
Figure 6: Detection Error Trade-off curves for various α/β proportions in different protocols.
Figure 7: Detection Error Trade-off curves for various bona fide images selection in different protocols.
Table 1: APCER@BPCER = (0.1, 0.01) of our method for various α/β proportions and bona fide images selection in different
protocols. Org. and SM. correspond to cases where only original and selfmorph images respectively were chosen as bonafide
samples.
Method
APCER@BPCER = δ
real facemorpher webmorph stylegan
δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01
α = 0 0.697 0.947 0.756 0.939 0.976 0.995 0.980 0.998
α/β = 0.01 0.601 0.895 0.651 0.957 0.945 0.992 0.895 0.991
α/β = 0.05 0.607 0.965 0.502 0.968 0.915 0.999 0.842 0.996
α/β = 0.2 0.401 0.835 0.431 0.786 0.778 0.979 0.799 0.969
α/β = 1.0 0.711 0.935 0.670 0.928 0.942 0.995 0.913 0.982
α/β = 10.0 0.556 0.839 0.513 0.791 0.749 0.923 0.852 0.969
β = 0.0 0.945 0.996 0.922 0.993 0.954 0.997 0.949 0.997
α/β = 0.2 Org. 0.481 0.875 0.498 0.876 0.987 0.997 0.915 0.993
α/β = 0.2 SM. 0.608 0.938 0.568 0.911 0.704 0.986 0.816 0.983
Table 2: BPCER@APCER = (0.1, 0.01) of our method for various α/β proportions and bona fide images selection in different
protocols. Org. and SM. correspond to cases where only original and selfmorph images respectively were chosen as bonafide
samples.
Method
BPCER@APCER = δ
real facemorpher webmorph stylegan
δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01 δ = 0.1 δ = 0.01
α = 0 0.795 0.993 0.781 0.971 0.961 0.998 0.952 0.998
α/β = 0.01 0.625 0.968 0.638 0.979 0.969 0.997 0.991 1.0
α/β = 0.05 0.577 0.950 0.598 0.951 0.841 0.989 0.895 0.991
α/β = 0.2 0.494 0.726 0.562 0.848 0.710 0.822 0.697 0.904
α/β = 1.0 0.616 0.871 0.628 0.843 0.882 0.963 0.851 0.960
α/β = 10.0 0.642 0.892 0.604 0.913 0.742 0.931 0.829 0.967
β = 0.0 0.956 0.998 0.932 0.998 0.971 0.998 0.874 0.986
α/β = 0.2 Org. 0.417 0.601 0.370 0.595 0.718 0.963 0.605 0.862
α/β = 0.2 SM. 0.580 0.912 0.587 0.905 0.484 0.842 0.798 0.976
MorDeephy: Face Morphing Detection via Fused Classification
199
protocol-facemorpher (2k morphs), which in-
cludes simple morphs with foreground and back-
ground artifacts
protocol-webmorph (1k morphs), which in-
cludes images with background artifacts but the
low level of artifacts inside the face contour
protocol-stylegan (2k morphs), which includes
StyleGan morphs
As bona fide images all our protocols use frontal
faces from the following public datasets: FRLL
Set(DeBruine and Jones, 2017), FEI (Artificial In-
telligence Laboratory of FEI in S
˜
ao Bernardo do
Campo, S
˜
ao Paulo, Brazil, 2006), AR(Martinez and
Benavente, 1998), Aberdeen and Utrecht (School of
Natural Sciences University of Stirling, 1998) (1.5k
images in total).
Since our benchmark utilities are not public yet
other developers cannot contribute their results to per-
form SOTA comparison for this work. That is why
we use our protocols only for performing the ablation
study and finding the best settings for our approach.
6 EXPERIMENTS AND RESULTS
To analyze the performance of our approach we
make several experiments with our method. As
backbone networks we use ResNet-50 (He et al.,
2016), which are initialized with weights, pretrained
on the ImageNet dataset. Followed by pooling
and dropout layers each backbone returns 512 deep
features. To complete the networks for classifica-
tion we add the classification layer with the size
= C equal to the number of classes in the train-
ing dataset. Input images (RGB 3-channel) are
aligned and resized to 224×224. In all our exper-
iments we perform training with SGD (Stochastic
Gradient Descent) optimizer for 2 epochs, decreas-
ing learning rate from 0.001 to 0.0001. We report the
performance by APCER@BPCER = (0.1,0.01) and
BPCER@APCER = (0.1,0.01).
Our default training dataset is a joined and shuf-
fled concatenation of the above mentioned datasets:
VGGFace2-selected, VGGFace2-selected-selfmorph,
and VGGFace2-selected-morph.
It is important to note that in all experiments we
assured the equal balance between the numbers of
morphed and non-morphed (which are bona fide +
selfmorphed) images in the training dataset.
6.1 Fused Classification Balance
For effective convergence and further morphing de-
tection, our method requires choosing the proper
balance between the elements of the loss function.
Namely the balance between α (= α
1
= α
2
) and β
(disbalance of α
1
and α
2
didn’t demonstrate any in-
teresting behaviour in our tests). We perform train-
ing of our method with different proportional settings
also including the ablation of particular parts from the
overall loss. Our experiments demonstrate (see Fig.
6 and Tab. 1, 2) that by varying α/β proportion it
is possible to achieve some optimal performance of
morphing detection in different protocols. Our strat-
egy allows generalizing the detection of morphing
even to the images, which are generated with GANs
even accounting that this type of morphing is totally
unseen in the training.
On the edge case with excluded main loss func-
tion driver (namely binary morph/bona fide classifi-
cation), our method demonstrates the almost random
detection decision. At the same time, ablation of the
regularization (β = 0) also leads to bad performance,
which we relate with the overfitting on the trivial bi-
nary classification learning task.
Summing up, we can conclude that our strategy
allows learning such face features which are discrim-
inative by the criteria of authenticity.
6.2 Data Combination Experiments
Further experiments are performed with the selected
proportion α/β = 0.2 in order to understand the im-
pact of selfmorphing for our method.
In comparison to the dataset selection in Section
6.1, where the collection of bona fide samples is split
evenly to original and selfmorphs, we test two more
options where these particular parts are ablated from
the total dataset.
Our results (see Fig. 7 and Tab. 1,2) proves the
significant importance of selfmorphs in our strategy.
Utilizing selfmorphs at the training stage allows to re-
duce the emphasis of the detection of facial blending
artefacts and shift it to the behavior of the deep feature
for generalizing to unseen types of attacks.
6.3 NIST FRVT Morph Results
We have performed the comparison of the results of
our method and the state of the art face morphing de-
tection approaches with NIST FRVT MORPH Bench-
mark (NIST, 2022).
We select the model with α/β = 0.2 from Section
6.1 as our best model and present results of compari-
son in several protocols:
P1 - Visa-Border (25727 Morphs)
P2 - Manual (323 Morphs)
P3 - MIPGAN-II (2464 Morphs)
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
200
Table 3: Comparison with the state of the art single image morphing detection methods by APCER@BPCER = (0.1, 0.01).
Method
APCER@BPCER = δ
P1 P2 P3 P4 P5
δ = δ = δ = δ = δ = δ = δ = δ = δ = δ =
0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01
(Aghdaie et al., 2021) 0.172 0.782 0.965 0.998 0.923 0.991 0.015 0.200 0.271 0.721
(Debiasi et al., 2018) 0.994 0.996 0.049 0.823 0.994 1.000 1.000 1.000 0.985 0.994
(Ramachandra et al., 2019) 0.659 0.996 0.375 0.990 0.938 0.985 0.159 0.998 0.936 0.996
(Scherhag et al., 2018) 0.986 1.000 1.000 1.000 0.997 1.000 0.996 1.000 0.993 1.000
(Lorenz et al., 2021) 0.844 1.000 0.380 1.000 0.966 1.000 0.819 1.000 0.971 0.995
(Ferrara et al., 2021) 0.982 0.996 0.477 0.999 0.978 1.000 0.037 0.810 0.420 0.777
Ours 0.419 0.658 0.434 0.686 0.842 0.954 0.323 0.639 0.499 0.805
P4 - Print + Scanned (3604 Morphs)
As bona fide samples all protocols utilize
a large collection of 1047389 Bona Fide im-
ages. The comparison is performed by the metrics
APCER@BPCER = (0.1, 0.01).
6.4 Single Image MAD
We perform the comparison in the target single image
morphing detection scenario (see Table 3).
MorDeephy outperforms other techniques in de-
tecting landmark-based morphs and challenging man-
ual morphs and achieves comparable results in other
protocols.
Also, our method does not demonstrate bias to a
particular morphing generative strategy and has the
most stable performance across all protocols in com-
parison to other approaches.
It is important to note that these results are
achieved by utilizing a rather straightforward and sim-
ple morphing technique during training (without any
adaptation to realistic scenario or modifications for re-
moving artefacts), which proves that our method al-
lows generalizing morphing detection to various un-
seen generative approaches by focusing on deep face
features behavior.
7 CONCLUSION
We introduce a novel deep learning strategy for single
image face morphing detection, which implies utiliz-
ing a complex classification task. It is directed onto
learning the deep facial features, which carry infor-
mation about the authenticity of these features. Our
method achieved the state of the art performance and
demonstrated a prominent ability for generalizing the
task of morphing detection to unseen scenarios (like
GAN morphs and print/scan morphs).
Our work also introduces several additional con-
tributions, which are the public and easy-to-use face
morphing detection benchmark and the results of our
wild datasets filtering strategy.
In our further work, we will focus on improv-
ing the performance by applying more sophisticated
morphing techniques during training and on explicit
adapting our method to the differential scenario,
which will require sophisticated sampling strategies.
ACKNOWLEDGEMENTS
The authors would like to thank the Portuguese Mint
and Official Printing Office (INCM) and the Institute
of Systems and Robotics - the University of Coimbra
for the support of the project Facing. This work has
been supported by Fundac¸
˜
ao para a Ci
ˆ
encia e a Tec-
nologia (FCT) under the project UIDB/00048/2020.
The computational part of this work was performed
with the support of NVIDIA Applied Research Accel-
erator Program with hardware and software provided
by NVIDIA.
REFERENCES
Aghdaie, P., Chaudhary, B., Soleymani, S., Dawson, J. M.,
and Nasrabadi, N. M. (2021). Attention Aware
Wavelet-based Detection of Morphed Face Images.
2021 IEEE International Joint Conference on Biomet-
rics (IJCB), pages 1–8.
An, X., Zhu, X., Gao, Y., Xiao, Y., Zhao, Y., Feng, Z., Wu,
L., Qin, B., Zhang, M., Zhang, D., and Fu, Y. (2021).
Partial FC: Training 10 Million Identities on a Single
Machine. In Proceedings of the IEEE/CVF Interna-
tional Conference on Computer Vision (ICCV) Work-
shops, pages 1445–1449.
Artificial Intelligence Laboratory of FEI in S
˜
ao Bernardo
do Campo, S
˜
ao Paulo, Brazil (2006). FEI face
database. https://fei.edu.br/ cet/facedatabase.html.
(accessed: September 1, 2022).
MorDeephy: Face Morphing Detection via Fused Classification
201
Bansal, R., Raj, G., and Choudhury, T. (2016). Blur im-
age detection using laplacian operator and open-cv. In
2016 International Conference System Modeling Ad-
vancement in Research Trends (SMART), pages 63–
67.
Borghi, G., Franco, A., Graffieti, G., and Maltoni, D.
(2021a). Automated artifact retouching in morphed
images with attention maps. IEEE Access, 9:136561–
136579.
Borghi, G., Pancisi, E., Ferrara, M., and Maltoni, D.
(2021b). A double siamese framework for differen-
tial morphing attack detection. Sensors, 21:3466.
Cao, Q., Shen, L., Xie, W., Parkhi, O. M., and Zisserman,
A. (2018). Vggface2: A dataset for recognising faces
across pose and age. In International Conference on
Automatic Face and Gesture Recognition.
Chopra, S., Hadsell, R., and LeCun, Y. (2005). Learning a
similarity metric discriminatively, with application to
face verification. 2005 IEEE Computer Society Con-
ference on Computer Vision and Pattern Recognition
(CVPR’05), 1:539–546 vol. 1.
Damer, N., Saladi
´
e, A. M., Braun, A., and Kuijper, A.
(2018). MorGAN: Recognition Vulnerability and At-
tack Detectability of Face Morphing Attacks Created
by Generative Adversarial Network. In 2018 IEEE 9th
International Conference on Biometrics Theory, Ap-
plications and Systems (BTAS), pages 1–10.
Damer, N., Saladi
´
e, A. M., Zienert, S., Wainakh, Y.,
Terh
¨
orst, P., Kirchbuchner, F., and Kuijper, A. (2019).
To Detect or not to Detect: The Right Faces to
Morph. In 2019 International Conference on Biomet-
rics (ICB), pages 1–8.
Damer, N., Spiller, N., Fang, M., Boutros, F., Kirchbuchner,
F., and Kuijper, A. (2021). PW-MAD: Pixel-wise Su-
pervision for Generalized Face Morphing Attack De-
tection. ArXiv, abs/2108.10291.
Debiasi, L., Scherhag, U., Rathgeb, C., Uhl, A., and Busch,
C. (2018). PRNU-based detection of morphed face
images. 2018 International Workshop on Biometrics
and Forensics (IWBF), pages 1–7.
DeBruine, L. and Jones, B. (2017).
Face research lab london set.
https://figshare.com/articles/dataset/Face Research Lab
London Set/5047666/3. (accessed: September 1,
2022).
Deng, J., Guo, J., Xue, N., and Zafeiriou, S. (2019). Ar-
cFace: Additive Angular Margin Loss for Deep Face
Recognition. In 2019 IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR), pages
4685–4694.
Dorizzi, B., Cappelli, R., Ferrara, M., Maio, D., Maltoni,
D., Houmani, N., Garcia-Salicetti, S., and Mayoue,
A. (2009). Fingerprint and On-Line Signature Verifi-
cation Competitions at ICB 2009. In Advances in Bio-
metrics : Third International Conference, ICB 2009,
Alghero, Italy, volume 5558.
Dustone, T. (2017). New face morph-
ing database for vulnerability research.
https://www.linkedin.com/pulse/new-face-morphing-
dataset-vulnerability-research-ted-dunstone. (ac-
cessed: September 1, 2022).
Ferrara, M., Franco, A., and Maltoni, D. (2014). The magic
passport. IJCB 2014 - 2014 IEEE/IAPR International
Joint Conference on Biometrics.
Ferrara, M., Franco, A., and Maltoni, D. (2018). Face de-
morphing. IEEE Transactions on Information Foren-
sics and Security, 13(4):1008–1017.
Ferrara, M., Franco, A., and Maltoni, D. (2021). Face mor-
phing detection in the presence of printing/scanning
and heterogeneous image sources. IET Biometrics, 10.
Guo, Y., Zhang, L., Hu, Y., He, X., and Gao, J. (2016). MS-
Celeb-1M: A Dataset and Benchmark for Large-Scale
Face Recognition. In Proceedings of ECCV, volume
9907, pages 87–102.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In 2016 IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR), pages 770–778, Las Vegas, NV, USA. IEEE.
Hernandez-Ortega, J., Galbally, J., Fierrez, J., Haraksim,
R., and Beslay, L. (2019). FaceQnet: Quality As-
sessment for Face Recognition based on Deep Learn-
ing. In 2019 International Conference on Biometrics
(ICB), pages 1–8.
Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J.,
and Huang, F. (2020). CurricularFace: Adaptive Cur-
riculum Learning Loss for Deep Face Recognition. In
The IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR).
International Organization for Standardization (09 2017).
ISO/IEC 30107–3:2017. Information Technol-
ogy—Biometric Presentation Attack Detection
Part 3: Testing and Reporting. ISO/IEC JTC 1/SC 37
Biometrics.
Jin, C., Jin, R., Chen, K., and Dou, Y. (2018). A com-
munity detection approach to cleaning extremely large
face database. Computational intelligence and neuro-
science, 2018.
Laboratory, B. S. (2018). UBO-Morpher.
http://biolab.csr.unibo.it/Research.asp. (accessed:
September 1, 2022).
Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., and Song, L.
(2017). SphereFace: Deep Hypersphere Embedding
for Face Recognition. In 2017 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR),
pages 6738–6746.
LLC, M. M. (2010). FaceFusion application.
www.wearemoment.com/FaceFusion/. (accessed:
September 1, 2022).
Lorenz, S., Scherhag, U., Rathgeb, C., and Busch., C.
(2021). Morphing attack detection: A fusion ap-
proach. In IEEE Fusion.
Martinez, A. and Benavente, R. (1998). The AR face
database. Tech. Rep. 24 CVC Technical Report.
Medvedev, I., Gonc¸alves, N., and Cruz, L. (2020). Biomet-
ric System for Mobile Validation of ID And Travel
Documents. In 2020 International Conference of the
Biometrics Special Interest Group (BIOSIG), pages 1–
5.
Mittal, A., Moorthy, A. K., and Bovik, A. C. (2012).
No-reference image quality assessment in the spatial
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
202
domain. IEEE Transactions on Image Processing,
21(12):4695–4708.
Neubert, T., Kraetzer, C., and Dittmann, J. (2019). A Face
Morphing Detection Concept with a Frequency and a
Spatial Domain Feature Space for Images on eMRTD.
In Proceedings of the ACM Workshop, pages 95–100.
Neubert, T., Makrushin, A., Hildebrandt, M., Kraetzer, C.,
and Dittmann, J. (2018). Extended stirtrace bench-
marking of biometric and forensic qualities of mor-
phed face images. IET Biometrics, 7(4):325–332.
NIST (2020). International Face Performance
Conference. https://www.nist.gov/news-
events/events/2020/10/international-face-
performance-conference-ifpc-2020. (accessed:
September 1, 2022).
NIST (2022). NIST FRVT MORPH.
https://pages.nist.gov/frvt/html/frvt morph.html.
(accessed: September 1, 2022).
Phillips, P. J., Flynn, P., Scruggs, T., Bowyer, K., Chang,
J. K., Hoffman, K., Marques, J., Min, J., and Worek,
W. (2005). Overview of the Face Recognition Grand
Challenge. In Proceedings of IEEE Computer Society
Conference on Computer Vision and Pattern Recogni-
tion, volume 1, pages 947– 954.
Raghavendra, R., Raja, K. B., and Busch, C. (2016). Detect-
ing morphed face images. In 2016 IEEE 8th Interna-
tional Conference on Biometrics Theory, Applications
and Systems (BTAS), pages 1–7.
Raghavendra, R., Raja, K. B., Venkatesh, S., and Busch, C.
(2017). Transferable Deep-CNN Features for Detect-
ing Digital and Print-Scanned Morphed Face Images.
2017 IEEE Conference on Computer Vision and Pat-
tern Recognition Workshops (CVPRW), pages 1822–
1830.
Raja, K., Ferrara, M., Franco, A., Spreeuwers, L., Batskos,
I., Wit, F., Gomez-Barrero, M., Scherhag, U., Fischer,
D., Venkatesh, S., Singh, J. M., Li, G., Bergeron, L.,
Isadskiy, S., Raghavendra, R., Rathgeb, C., Frings, D.,
Seidel, U., Knopjes, F., and Busch, C. (2020). Morph-
ing Attack Detection - Database, Evaluation Platform
and Benchmarking. IEEE Transactions on Informa-
tion Forensics and Security, PP:1–1.
Ramachandra, R., Venkatesh, S., Raja, K., and Busch, C.
(2019). Towards making morphing attack detection
robust using hybrid scale-space colour texture fea-
tures. In 2019 IEEE 5th International Conference
on Identity, Security, and Behavior Analysis (ISBA),
pages 1–8.
Research, T. C. and Service, D. I. (2020). Im-
age Manipulation Attack Resolving Solutions.
https://cordis.europa.eu/project/id/883356. (accessed:
September 1, 2022).
Ruiz, N., Chong, E., and Rehg, J. (2018). Fine-grained head
pose estimation without keypoints. In The IEEE Con-
ference on Conference on Computer Vision and Pat-
tern Recognition (CVPR) Workshops.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,
S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).
ImageNet Large Scale Visual Recognition Challenge.
International Journal of Computer Vision (IJCV),
115(3):211–252.
Sarkar, E., Korshunov, P., Colbois, L., and Marcel, S.
(2020). Vulnerability analysis of face morphing at-
tacks from landmarks and generative adversarial net-
works. arXiv preprint.
Satya, M. (2016). Face Morph Using OpenCV.
www.learnopencv.com/face-morph-using-opencv-
cpp-python/. (accessed: September 1, 2022).
Scherhag, U., Debiasi, L., Rathgeb, C., Busch, C., and Uhl,
A. (2019). Detection of Face Morphing Attacks Based
on PRNU Analysis. IEEE Transactions on Biomet-
rics, Behavior, and Identity Science, 1(4):302–317.
Scherhag, U., Raghavendra, R., Raja, K. B., Gomez-
Barrero, M., Rathgeb, C., and Busch, C. (2017). On
the vulnerability of face recognition systems towards
morphed face attacks. In 2017 5th International Work-
shop on Biometrics and Forensics (IWBF), pages 1–6.
Scherhag, U., Rathgeb, C., and Busch, C. (2018).
Morph Deterction from Single Face Image: a Multi-
Algorithm Fusion Approach. In Proceedings of the
2018 2nd International Conference on Biometric En-
gineering and Applications, pages 6–12.
Scherhag, U., Rathgeb, C., Merkle, J., and Busch, C.
(2020). Deep Face Representations for Differential
Morphing Attack Detection. IEEE Transactions on
Information Forensics and Security, 15:3625–3639.
School of Natural Sciences University of Stirling
(1998). Psychological Image Collection of Stir-
ling. http://pics.stir.ac.uk. (accessed: September 1,
2022).
Schroff, F., Kalenichenko, D., and Philbin, J. (2015).
FaceNet: A unified embedding for face recognition
and clustering. In 2015 IEEE Conference CVPR,
pages 815–823.
Seibold, C., Samek, W., Hilsmann, A., and Eisert, P. (2017).
Detection of Face Morphing Attacks by Deep Learn-
ing. In Proceedings of Digital Forensics and Water-
marking: 16th International Workshop, IWDW 2017,
pages 107–120.
Shi, Y. and Jain, A. (2019). Probabilistic Face Embeddings.
In 2019 IEEE/CVF International Conference on Com-
puter Vision (ICCV), pages 6901–6910.
Shi, Y., Yu, X., Sohn, K., Chandraker, M., and Jain, A.
(2020). Towards Universal Representation Learning
for Deep Face Recognition. In Proceedings of 2020
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition (CVPR), pages 6816–6825.
Sun, J., Yang, W., Xue, J., and Liao, Q. (2020). An Equal-
ized Margin Loss for Face Recognition. IEEE Trans-
actions on Multimedia, pages 1–1.
Sun, Y., Chen, Y., Wang, X., and Tang, X. (2014). Deep
Learning Face Representation by Joint Identification-
Verification. In NIPS.
Sun, Y., Wang, X., and Tang, X. (2014). Deep Learning
Face Representation from Predicting 10,000 Classes.
In 2014 IEEE Conference on Computer Vision and
Pattern Recognition, pages 1891–1898.
Sun, Y., Wang, X., and Tang, X. (2015). Deeply learned
face representations are sparse, selective, and robust.
MorDeephy: Face Morphing Detection via Fused Classification
203
2015 IEEE Conference on Computer Vision and Pat-
tern Recognition (CVPR), pages 2892–2900.
Tremoc¸o, J., Medvedev, I., and Gonc¸alves, N. (2021). Qual-
Face: Adapting Deep Learning Face Recognition for
ID and Travel Documents with Quality Assessment.
In 2021 International Conference of the Biometrics
Special Interest Group (BIOSIG), pages 1–6.
Venkatesh, S. K., Zhang, H., Ramachandra, R., Raja, K. B.,
Damer, N., and Busch, C. (2020). Can GAN Gen-
erated Morphs Threaten Face Recognition Systems
Equally as Landmark Based Morphs? - Vulnerabil-
ity and Detection. 2020 8th International Workshop
on Biometrics and Forensics (IWBF), pages 1–6.
Wandzik, L., Kaeding, G., and Garcia, R. V. (2018). Morph-
ing Detection Using a General- Purpose Face Recog-
nition System. In 2018 26th European Signal Process-
ing Conference (EUSIPCO), pages 1012–1016.
Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J.,
Li, Z., and Liu, W. (2018). CosFace: Large Mar-
gin Cosine Loss for Deep Face Recognition. In 2018
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition, pages 5265–5274.
Wen, Y., Zhang, K., Li, Z., and Qiao, Y. (2016). A Dis-
criminative Feature Learning Approach for Deep Face
Recognition. In Leibe, B., Matas, J., Sebe, N., and
Welling, M., editors, Computer Vision ECCV 2016,
pages 499–515, Cham. Springer International Pub-
lishing.
Yi, D., Lei, Z., Liao, S., and Li, S. (2014). Learning face
representation from scratch. ArXiv, abs/1411.7923.
Zeng, D., Shi, H., Du, H., Wang, J., Lei, Z., and Mei, T.
(2020). NPCFace: A Negative-Positive Cooperation
Supervision for Training Large-scale Face Recogni-
tion. CoRR, abs/2007.10172.
Zhang, H., Venkatesh, S., Ramachandra, R., Raja, K.,
Damer, N., and Busch, C. (2020). MIPGAN - Gener-
ating Robust and High Quality Morph Attacks Using
Identity Prior Driven GAN. ArXiv, abs/2009.01729.
Zhang, L., Zhang, L., and Li, L. (2017). Illumination Qual-
ity Assessment for Face Images: A Benchmark and a
Convolutional Neural Networks Based Model. Lec-
ture Notes in Computer Science, 10636 LNCS:583–
593.
Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X.,
Zhu, J., Yang, T., Lu, J., Du, D., and Zhou, J. (2021).
WebFace260M: A Benchmark Unveiling the Power of
Million-Scale Deep Face Recognition. In Proceedings
of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), pages 10492–10502.
ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods
204