Does Melania Trump Have a Body Double from the Perspective of

Automatic Face Veriﬁcation?

Khawla Mallat

1 a

, Fabiola Becerra-Riera

, Annette Morales-Gonz

alez

2 b

, Heydi M

endez-V

azquez

and Jean-Luc Dugelay

1 c

Digital Security Department, EURECOM Campus Sophia Tech, 450 route des Chappes F-06410 Sophia Antipolis, France

Advanced Technologies Application Center (CENATAV) 7A #21406 Siboney, Playa, P. C. 12200, Havana, Cuba

Keywords:

Information Veriﬁcation, Automatic Face Veriﬁcation, Conspiracy Theory, Image Quality.

Abstract:

With the growing number of users getting updated about current events through social media, the spread of

misinformation is increasing and thus endorsing conspiracy belief. During the last presidential election cam-

paign in the USA, the conspiracy theory claiming the existence of a body double that stands in for the former

ﬁrst lady Melania Trump had made international news headlines. Fighting the spread of misinformation is

crucial as it is threatening the society by manipulating the public opinion. In this paper, we explore whether

automatic face veriﬁcation can help in verifying widespread misinformation on social media, dealing partic-

ularly with the conspiracy theory related to Melania Trump replacement. We employed four different state-

of-the-art descriptors for face recognition to verify the integrity of the claim of the studied conspiracy theory.

In addition, we assessed the impact of different image quality metrics on the variation of the face veriﬁcation

scores. Two sets of image quality metrics were considered: acquisition-related metrics and subject-related

metrics.

1 INTRODUCTION

This is not Melania! In October 2017, the hypothesis

of the existence of a body double of Melania TRUMP

appeared for the ﬁrst time on social networks. Then

regularly, what is now referred to as ‘The Melania

Trump replacement conspiracy theory’

by mass me-

dia, reappears in the front of the stage, at least once a

year, with a peak during the last presidential election

campaign. Face is used by all of us as the main human

trait to recognize each other in daily life. This is why,

even if some other criteria have been used to demon-

strate that Melania is not Melania, like her attitude

and behavior or her body height, most arguments ad-

vanced by people are related to her facial appearance.

Algorithms based on artiﬁcial intelligence, im-

age/video processing and computer vision are de-

signed and used to ﬁght against fake news that in-

clude doctored images (Mahfoudi et al., 2019) and/or

AI-based generated videos (i.e. deepfakes) (Dolhan-

https://orcid.org/0000-0003-2116-8219

https://orcid.org/0000-0003-2716-3144

https://orcid.org/0000-0003-3151-4330

https://w.wiki/4MVP

sky et al., 2020). The issue addressed in this article

is different in the sense that the open question is not

about the integrity of the picture but about the physi-

cal presence of the real Melania or of a body double in

the scene recorded by the camera. So Biometrics (and

more particularly Automatic Face Recognition), more

than image forensic tools, is the appropriate technol-

ogy to address the problem.

There are two possibilities in case of the non-

existence of a body double of Melania: either peo-

ple are complotists in the sense that they are aware

that Melania has no double body but claim the con-

trary or, people are of good faith and consider by vi-

sual inspection that there are two different people on

photographs: Melania and a body double based on

some possible visual differences among faces. Visual

inspection by human is subjective and today less ac-

curate than a digital and objective inspection by ma-

chines for many tasks related to face analysis (Phillips

and O’Toole, 2014). This is why automatic face

veriﬁcation may signiﬁcantly help in the process for

forming a robust opinion concerning this supposed

conspiracy theory.

In this work, we use automatic face veriﬁcation to

conduct an analysis of a group of images of Melania

Mallat, K., Becerra-Riera, F., Morales-González, A., Méndez-Vázquez, H. and Dugelay, J.

Does Melania Trump Have a Body Double from the Perspective of Automatic Face Veriﬁcation?.

DOI: 10.5220/0010911100003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 737-744

ISBN: 978-989-758-549-4; ISSN: 2184-4313

737

Trump, in order to try to determine if she has a body

double or not.

2 AUTOMATIC FACE

VERIFICATION

In this section, we present primarily the image sam-

ples of Melania Trump on which our study has been

based. Then, we introduce the state-of-the-art de-

scriptors of face recognition employed in this study

followed by the experimental protocol and results

that will help answering the question ‘Does Melania

Trump has a body double?’.

2.1 Face Samples

The presented study was driven by the interest of the

French metropole television M6 in the conspiracy the-

ory case related to Melania Trump

. The TV channel

provided us with a set of images of Melania Trump

collected from social media, mainly Twitter, collected

by a team of journalists working on the topic. These

images were labeled by the social media users as the

‘real’ Melania Trump or as the ‘body double’. The to-

tal number of the image samples of Melania Trump is

30. Considering the labels provided by the journalists,

16 images were labeled as the ‘real’ Melania Trump

and 14 as the ‘body double’ . Faces were detected us-

ing RetinaFace (Deng et al., 2020) and only faces of

Melania Trump were selected and then aligned. Fig-

ure 1 shows the preprocessed face images. At ﬁrst

glance, one could see that there is an important dis-

crepancy between the two sets of images. First, the set

of the alleged ‘body double’ contains 6 out of 14 im-

ages in which Melania is wearing large dark glasses

which covers up the ocular area, while it is the case

for only 1 out of 16 images for the ‘real’ Melania.

In addition, one could see that the set of images of

the ‘real’ Melania contains images where Melania is

captured with a smiling face, open eyes and proper il-

lumination conditions, while the images contained in

the set of the alleged ‘body double’ of Melania seem

to be taken from video recordings, where Melania ap-

pears with blinking eyes, head down and low illumi-

nation conditions.

2.2 Face Recognition Models

In this study, we employ four state-of-the-art

face recognition descriptors that achieved more

than 99.3% of accuracy when evaluated on LFW

https://youtu.be/9HJKKYeJdTk?t=1239

dataset (Huang et al., 2007). For each face descrip-

tor, feature embeddings are extracted from Melania’s

face images and compared against each other using

cosine similarity to perform the face veriﬁcation task.

LightCNN (Wu et al., 2018). is an implementation

of CNN for face recognition designed to have fewer

trainable parameters than a vanilla CNN and to han-

dle noisy labels. This network introduces a new con-

cept of max-out activation in each convolutional layer

for feature ﬁlter selection. We use the pretrained net-

work with 29-layers on LFW dataset (Huang et al.,

2007) to obtain embeddings of 256-dimension from

face images.

MobileSqueezeNet (mob, 2019). is a lightweight

CNN released with Intel’s OpenVino framework and

model zoo, for face recognition in re-identiﬁcation

scenarios. It is based on MobileNet V2 (Howard

et al., 2017) backbone, which consists of 3x3 in-

verted residual blocks with squeeze-excitation atten-

tion modules. After the backbone, the network ap-

plies global depthwise pooling and then uses 1x1

convolution to create the ﬁnal 256-dimensional em-

bedding vector. We used the pretrained model

face-reidentification-retail-0095.

SphereFace (Liu et al., 2017). refers to angu-

larly distributed feature embeddings designed for face

recognition. In contrast to employing Euclidean mar-

gin to the CNN’s learned features, SphereFace pro-

poses to use an angular softmax (named A-Softmax)

to learn discriminative face features with a novel ge-

ometric interpretation. A-Softmax loss can be in-

terpreted as constraining learned features to be dis-

criminative on a hypersphere manifold, which intrin-

sically matches the prior that face images lie on a

manifold. We use the 512-dimensional SphereFace

embeddings extracted using the pretrained model on

CASIA-WebFace dataset (Yi et al., 2014).

ArcFace (Deng et al., 2019). is Additive Angular

Margin loss function, a new loss function for face

recognition designed to improve the discriminative

power of the learnt feature embeddings. The pro-

posed loss function aims to optimise the geodesic

distance margin by considering the correspondence

between the angle and arc in the normalised hyper-

sphere. We used a model pretrained on CASIA-

WebFace dataset (Yi et al., 2014) to extract embed-

dings of 512-dimension from face images.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

738

Image samples of the

real ”

Melania Trump

Image samples of the alleged

body double ”

of Melania Trump

Figure 1: Preprocessed face images of the ‘real’ Melania and the alleged ‘body double’.

(a) (b)

Genuines (Melania) Imposters (Melania) Genuines (Celebrities) Imposters( Celebrities)

Figure 2: Similarity score distribution of ‘real’ against ‘body double’ Melania Trump (a) Arcface (b) LightCNN (c) Mobi-

leSqueezeNet (d) SphereFace.

2.3 Experimental Protocol and Results

Assuming we have 2 identities for Melania Trump,

the ﬁrst consists of the ‘real’ Melania and the second

consists of the alleged ‘body double’, we performed a

face veriﬁcation experiment using an All-vs-All com-

parison scheme. For each identity, we performed the

face veriﬁcation experiment by considering a differ-

ent reference face image at a time, each reference face

image was compared against all the remaining image

samples of the 2 identities, resulting in 224 genuine

trials and 422 imposter trials. Genuine pairs belong

to the same identity, while Impostors pairs correspond

to the comparison of one identity vs. the other, mean-

ing the ‘real’ Melania against the alleged ‘body dou-

ble’. We illustrate in Figure 2 the score distributions

of Genuines in green and of Impostors in red using the

four selected face descriptors. We also introduced the

score distributions of imposters and genuines deduced

from a dataset, collected from the web by the authors,

of female celebrities in their 50s similar in proﬁle to

Melania Trump. This dataset contains 50 identities,

with 30 images per identity.

It can be seen that the genuine (in green) and the

imposter (in red) score distributions overlap substan-

tially, although, there is a slight difference between

the two score distributions. This difference might be

due to the discrepancy in image samples between the

set of ‘real’ Melania and the alleged ‘body double’

as depicted in section 2.1. However, we observe that

the genuine and the imposter similarity score distribu-

tions deduced from the celebrities dataset are almost

completely disjoint, as the dataset contains different

identities. In section 3, we evaluate the impact of

different factors that might affect the automatic face

Does Melania Trump Have a Body Double from the Perspective of Automatic Face Veriﬁcation?

739

Table 1: Equal error rate (EER) and area under the curve (AUC) of the different face veriﬁcation systems.

ArcFace LightCNN MobileSqueezeNet SphereFace

EER (%) AUC (%) EER (%) AUC (%) EER (%) AUC (%) EER (%) AUC (%)

Melania 39.3 66.2 37 68 38.8 66.2 39.3 65.5

Celebrities 0.27 99.2 0.22 99.2 0.20 99.4 0.43 98.7

veriﬁcation score as an attempt to justify the slight

difference between the green and the red score dis-

tribution. In addition, one could see that the distri-

butions of genuine and imposter scores deduced from

the Melania data generally fall into the genuine score

distribution deduced from the celebrities dataset, with

a minor overlap with the imposters score distribution.

Table 1 reports the equal error rates (EER) and the

area under the curve (AUC) deduced from the face

veriﬁcation experiment performed on Melania image

samples and on the celebrities dataset. We note that

for distinct identities, as it is the case for the celebri-

ties dataset, 0.02% of EER and 0.99% of AUC are

reported in average. However, when computed for

the Melania face veriﬁcation experiment, the EER in-

creased to 0.38% and the AUC decreased to 0.66% in

average, implying that the 2 considered identities ( the

‘real’ Melania and the alleged ‘body double’ ) are ac-

tually the same. The results reported by the different

face recognition models, in ﬁgure 2 and table 1 are

highly comparable. The selection of 4 different state-

of-the-art face recognition models was motivated by

the need to validate our conclusions.

3 FACTORS AFFECTING

AUTOMATIC FACE

VERIFICATION

In this section, we suppose that Melania Trump does

not have a body double according to the signiﬁcant

overlap between genuine and imposter score distri-

butions deduced from Melania image set, illustrated

in section 2.3. However, in this section, we attempt

to justify the difference between these two distribu-

tions highlighted hereabove. There are several factors

that could affect the utility of a face image for a spe-

ciﬁc task. In this section, since we are particularly

interested in face veriﬁcation, we explored the im-

pact of key quality metrics (M

endez-V

azquez et al.,

2012), (Mallat et al., 2019) which were considered

according to the particular conditions of our case of

study and that are subject and environmental/sensor-

related.

To highlight the impact of a given quality met-

ric on the face veriﬁcation performance, we sorted all

the 30 image samples of Melania Trump according to

their corresponding quality measure in an ascending

order. Feature embeddings were extracted from all the

images. Then, the afﬁnity matrix of the cosine simi-

larity scores is computed between feature embeddings

of Melania Trump’s images in an All-vs-All compar-

ison scheme. Hereafter, only results obtained using

ArcFace feature embeddings will be reported as all

the four face descriptors depict the same behaviour.

3.1 Acquisition-related Image Quality

Metrics

From simply looking at the image samples of Melania

Trump, one could see a prominent difference between

the ‘real’ and the alleged ‘body double’ sets mainly in

the image acquisition quality.

Brightness. (Zohra et al., 2017) Depending on the

acquisition conditions and the camera devices, im-

ages could often be affected by variations in terms of

illumination and color leveling (the degree of light-

ness or darkness of a color) which result in bright-

ness variations. Brightness is the perception caused

by the luminance of an object, a face in our case. Ma-

jor ﬂuctuations of these measures could drastically

change the appearance of subjects causing different

skin color and shaded or extremely brighten zones

that could even partially occlude discriminative face

parts. Since brightness does not visually appear to be

homogeneous across the Melania Trump image set,

we decided to analyse how much it really affects the

performance of face veriﬁcation.

Face Luminance. (M

endez-V

azquez et al., 2012)

Face luminance deals with the distribution of light

across the face and it is measured in speciﬁc face re-

gions that are prominent for the face veriﬁcation pro-

cess. The face image quality based on luminance,

was measured by averaging the normalized mean lu-

minance values for a set of triangles, representing key

face regions, that are more likely to be affected by

changes in illumination. Since illumination does not

appear to be homogeneously distributed across the

face, an assessment of the variation of face luminance

on the face veriﬁcation score is conducted.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

740

(a) (b) (c) (d)

(e)

image labeled as ‘real’ Melania image labeled as ‘body double’

Figure 3: Afﬁnity matrix of cosine similarity between ArcFace embeddings of Melania Trump’s images, sorted according to:

(a) Brightness (b) Face luminance (c) Contrast (d) Exposure (e) Sharpness. Blue cells reﬂect a low similarity score (closer to

0) whereas green cells correspond to high similarity score (close to 1).

Exposure. (Shirvaikar, 2004) is the amount of light

that reaches the acquisition sensor and it can be con-

trolled using aperture and shutter settings. When the

face image is correctly exposed, the image provides

rich details in both dark shadow regions as well as

bright regions. The Melania Trump set of images con-

tain mainly underexposed images in which the dark

shadow regions lose ﬁne details and appear as pure

block regions because of receiving too little light.

Contrast. (Gao et al., 2007) Low contrast images

are commonly due to poor illumination conditions

during the image acquisition, wrong setting of the

aperture and shutter speed. Low contrast face images

contain underexposed and overexposed regions that

result in loss of geometrical and color information of

facial attributes. Image contrast is deﬁned as the sep-

aration between the darkest and brightest areas of the

image, and it can be computed as the difference in

color intensities.

Sharpness. (M

endez-V

azquez et al., 2012) In

blurred images, the face regions (eyes, nose, mouth,

etc.) are usually not sufﬁciently clear or well delim-

ited due to the loss of visual information by confus-

ing and mixing shapes, textures and colors. To decide

to which extent more or less deﬁned/sharp edges are

inﬂuencing the face veriﬁcation process, we applied

an unsharp masking process to sharpen the original

face image and average the resulting pixel intensities

to obtain a sharpness quality score.

3.2 Subject-related Image Quality

Metrics

The second main difference between the set of the

‘real’ and the alleged ‘body double’ is related to the

physical appearance of Melania herself. One could

see that the set of the ‘real’ Melania shows images of

her with the head straight and a glowy face, whilst for

the set of the alleged ‘body double’, Melania appears

either wearing sunglasses or with the eyes closed with

a varying head pose. Accordingly, 3 quality metrics

that are subject-related were selected to evaluate their

impact on face veriﬁcation similarity score.

Occlusion by Sunglasses. (M

endez-V

azquez

et al., 2012) Since several samples of the alleged

‘body double’ set show Melania wearing sunglasses,

our focus was drawn on the occlusion of the peri-

ocular region, which has been shown to be a very

discriminative region in face recognition, and its

impact on face veriﬁcation performance. Therefore,

to measure the quality of a face image based on the

use of sunglasses, we ﬁrst used a CNN classiﬁer as

sunglasses detector, and then we employed the output

probability of the class representing “non sunglasses”

as ﬁnal quality score: face images with higher ‘non

sunglasses’ probability are more likely to be good

quality images and vice versa.

Femininity. (Mallat et al., 2017) A main factor that

might have affected the face veriﬁcation performance

Does Melania Trump Have a Body Double from the Perspective of Automatic Face Veriﬁcation?

741

(a) (b) (c)

image labeled as ‘real’ Melania image labeled as ‘body double’

Figure 4: Afﬁnity matrix of cosine similarity between ArcFace embeddings of Melania Trump’s images, sorted according to:

(a) Occlusion by sunglasses (b) Femininity (c) Head pose. Blue cells reﬂect a low similarity score (closer to 0) whereas green

cells correspond to high similarity score (close to 1).

can clearly observed in the Melania Trump images is

her physical appearance. In the set of the ‘real’ Mela-

nia, she appears with radiant face and make up on,

whilst in the alleged ‘body double’ set, Melania seems

to be wearing too little make up or no make up at

all, with a wearied face. Consequently, we propose to

quantify the femininity appearance of Melania Trump

on all the image samples and evaluate their impact

on face veriﬁcation performance. For this purpose, a

gender estimation model (Antipov et al., 2016) was

employed to deliver a score representing the proba-

bility of being a male or a female. We consider the

femininity score as the probability of being a female.

Mallat et al. (Mallat et al., 2017) proved that the more

makeup is applied (up to a certain level), the higher

the femininity score is.

Head Pose. (M

endez-V

azquez et al., 2012) The

complexity of dealing with multiple head rotations

under uncontrolled conditions has long been a bottle-

neck in face recognition. Different head poses usually

imply the occlusion or deformation of some face re-

gions, which can lead to the mismatch of two samples

of the same person just because they were captured

from different angles. It is worth noting that general

Melania postures are basically frontal and just a few

samples present some yaw or pitch variation, in par-

ticular for the alleged “body double” set.

3.3 Discussion

Figure 3 reports the afﬁnity matrices of the cosine

similarity scores that highlight the impact of image

quality metrics that are related to the image acquisi-

tion: brightness, face luminance, sharpness, exposure

and contrast. Blue cells reﬂect a low similarity score

(closer to 0) whereas green cells correspond to high

similarity score (close to 1).

In Figures 3a, 3b, 3c and 3d, one could see that

images with low quality yield low similarity scores

and they correspond to image samples from the al-

leged ‘body double’, whilst high quality images de-

liver high similarity score and they are mostly asso-

ciated to image samples of ‘real’ Melania. For the

exposure quality metric, the lines of the afﬁnity ma-

trix, in Figure 3c, that depict low similarity scores for

images with higher exposure correspond to speciﬁc

image samples in which Melania Trump appears ei-

ther with sunglasses or with the head down. Figure 3e

shows a small correlation between sharper images and

high similarity scores. However, we can deduce from

the matrix that there are different and more relevant

factors that are inﬂuencing the face veriﬁcation per-

formance. The cross-shaped blue area that depict low

similarity scores from Figure 3e corresponds to the

comparison of four images of average sharpness but

considerably affected by illumination and using sun-

glasses; this is also the case of the two isolated sharp

images producing the blue lines that converge in the

bottom right corner of the matrix.

Figure 4 portrays the afﬁnity matrices of the co-

sine similarity scores by sorting the image samples

in an ascending order according to occlusion by sun-

glasses, femininity and head pose quality metrics, in

order to highlight their impact on face veriﬁcation

performance.

In Figures 4a and 4b, one could perceive that the

images with low quality scores yield low similarity

scores and correspond to the image samples of the

‘body double’, whereas the images with the highest

quality coincide with samples of the ‘real’ Melania

and they deliver relatively higher similarity scores.

The linear behaviour observed in Figure 4a is a conﬁr-

mation that sunglasses have high impact on the vari-

ation of face veriﬁcation scores when comparing the

set of ‘real’ Melania with the alleged ‘body double’.

The two crossed shaped blue areas, in Figure 4b, de-

picting low similarity scores for higher femininity val-

ues correspond to two samples of Melania wearing

sunglasses. However, Figure 4c shows the pose vari-

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

742

ation is barely inﬂuencing the face veriﬁcation per-

formance: comparisons between good and bad qual-

ity samples output higher or lower similarity scores

alike. This behaviour is an indicator that the used face

descriptors are robust enough to the slight head pose

variation across Melania’s image samples.

Among the selected quality metrics, one could af-

ﬁrm that brightness, face luminance and occlusion by

sunglasses are the most relevant quality metrics af-

fecting the face veriﬁcation performance, and to a

lesser degree with exposure, contrast and femininity.

Brightness, face luminance, exposure and contrast be-

ing all related to the illumination conditions during

the image acquisition depict the same behaviour of the

afﬁnity matrix that could be seen in ﬁgures 3a, 3b, 3c

and 3d. The slight shift between the genuine and

imposter score distributions deduced from Melania

Trump image samples, observed in subsection 2.3, is

certainly due to the fact that the set of the ‘real’ Mela-

nia contains overall images of better quality compared

to the set of the alleged ‘body double’, and obvi-

ously not because of the existence of a Melania Trump

‘body double’.

4 CONCLUSION

This paper proposes a ﬁrst attempt to explore whether

automatic face veriﬁcation can help in ﬁghting the

spread of conspiracy theories and if it could be used as

an information veriﬁcation tool. A further extensive

study should be performed on a wider set of identities

and their associated body doubles. In this paper, dif-

ferent state-of-the-art descriptors of face recognition

were used to compare the ‘real’ Melania image set to

the the images of the alleged ‘body double’. Accord-

ing to the proposed study, the answer to the question

‘Does Melania Trump have a body double?’ is deﬁ-

nitely a No. In the studied case, automatic face ver-

iﬁcation invalidates the polemical conspiracy theory

claiming the existence of a replacement of Melania

Trump during the presidency term of Donald Trump.

This disclaimer could only be reliable if we presume

that the provided face image samples were not priorly

manipulated. Recently, a new type of image manipu-

lation intended to trick the deep leaning based mod-

els into delivering an erroneous output. This type of

image manipulation, called adversarial attacks, is per-

formed by adding an adversarial perturbation that can

be imperceptible to the human eye (Bisogni et al.,

2021). It could be imagined, for instance, that a re-

placement of Melania Trump exists indeed and that

the associated image samples were manipulated in a

way that the replacement will be recognized as Mela-

nia Trump.

REFERENCES

(2019). Mobilesqueezenet: face reidentiﬁcation retail 0095.

t.ly/MqxX.

Antipov, G., Berrani, S.-A., and Dugelay, J.-L. (2016). Min-

imalistic cnn-based ensemble model for gender pre-

diction from face images. Pattern recognition letters.

Bisogni, C., Cascone, L., Dugelay, J.-L., and Pero, C.

(2021). Adversarial attacks through architectures and

spectra in face recognition. Pattern Recognition Let-

ters.

Deng, J., Guo, J., Ververas, E., Kotsia, I., and Zafeiriou, S.

(2020). Retinaface: Single-shot multi-level face local-

isation in the wild. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion.

Deng, J., Guo, J., Xue, N., and Zafeiriou, S. (2019). Ar-

cface: Additive angular margin loss for deep face

recognition. In Proceedings of the IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition,

pages 4690–4699.

Dolhansky, B., Howes, R., Pﬂaum, B., Baram, N., and Fer-

rer, C. C. (2020). The deepfake detection challenge

(dfdc) dataset. arXiv preprint arXiv:2006.07397.

Gao, X., Li, S. Z., Liu, R., and Zhang, P. (2007). Standard-

ization of face image sample quality. In International

Conference on Biometrics, pages 242–251. Springer.

Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D.,

Wang, W., Weyand, T., Andreetto, M., and Adam,

H. (2017). Mobilenets: Efﬁcient convolutional neu-

ral networks for mobile vision applications. arXiv

preprint arXiv:1704.04861.

Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller,

E. (2007). Labeled faces in the wild: A database for

studying face recognition in unconstrained environ-

ments. Technical Report 07-49, University of Mas-

sachusetts, Amherst.

Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., and Song, L.

(2017). Sphereface: Deep hypersphere embedding for

face recognition. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 212–220.

Mahfoudi, G., Tajini, B., Retraint, F., Morain-Nicolier, F.,

Dugelay, J. L., and Marc, P. (2019). Defacto: Image

and face manipulation dataset. In 27th European Sig-

nal Processing Conference (EUSIPCO).

Mallat, K., Damer, N., Boutros, F., and Dugelay, J.-L.

(2019). Robust face authentication based on dynamic

quality-weighted comparison of visible and thermal-

to-visible images to visible enrollments. In 2019 22th

International Conference on Information Fusion (FU-

SION), pages 1–8.

Mallat, K., Galdi, C., and Dugelay, J.-L. (2017). Evaluation

of the facial makeup impact on femininity appearance

based on automatic prediction. COMET 2017, 2nd

Cosmetic Measurement & Testing Symposium,

Cergy-Pontoise, France.

endez-V

azquez, H., Chang, L., Rizo-Rodr

ıguez, D., and

Morales-Gonz

alez, A. (2012). Evaluaci

on de la cali-

dad de las im

agenes de rostros utilizadas para la iden-

Does Melania Trump Have a Body Double from the Perspective of Automatic Face Veriﬁcation?

743

tiﬁcaci

on de las personas. Computaci

on y Sistemas,

16(2):147–165.

Phillips, P. J. and O’Toole, A. J. (2014). Comparison of

human and computer performance across face recog-

nition experiments. Image and Vision Computing,

32(1):74–85.

Shirvaikar, M. V. (2004). An optimal measure for camera

focus and exposure. In Thirty-Sixth Southeastern Sym-

posium on System Theory, 2004. Proceedings of the,

pages 472–475. IEEE.

Wu, X., He, R., Sun, Z., and Tan, T. (2018). A light cnn

for deep face representation with noisy labels. IEEE

Transactions on Information Forensics and Security.

Yi, D., Lei, Z., Liao, S., and Li, S. Z. (2014). Learn-

ing face representation from scratch. arXiv preprint

arXiv:1411.7923.

Zohra, F. T., Gavrilov, A. D., Duran, O. Z., and Gavrilova,

M. (2017). A linear regression model for estimat-

ing facial image quality. In 2017 IEEE 16th Inter-

national Conference on Cognitive Informatics & Cog-

nitive Computing (ICCI* CC), pages 130–138. IEEE.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

744