DLDFD: Recurrence Free 2D Convolution Approach for Deep Fake

Detection

Jag Mohan Singh and Raghavendra Ramachandra

Norwegian Biometrics Laboratory, Norwegian University of Science and Technology, Gjøvik, Norway

Keywords:

DeepFakes Detection, Resnet, Deep Learning Architecture.

Abstract:

Deep Fake images, which are digitally generated either through computer graphics or deep learning techniques,

pose an increasing risk to existing face recognition systems. This paper presents a Deep-Learning-based Deep

Fake Detection (DLDFD) architecture consisting of augmented convolutional layers followed by Resnet-50

architecture. We train DLDFD end-to-end with low-resolution images from the FaceForensics++ dataset. The

number of images used during different phases includes approximately 1.68 million during training, 315k

during validation, and 340k during testing. We train DLDFD in three different scenarios, combined image

manipulation where we achieve an accuracy of 96.07% compared to 85.14% of state-of-the-art (SOTA), single

image manipulation techniques where we get 100% accuracy for neural textures, and ﬁnally, cross-image

manipulation techniques where we achieve an accuracy of 94.28% on the unseen category of face swap much

higher than SOTA. Our approach requires only 2D convolutions without recurrence as compared to SOTA.

1 INTRODUCTION

Researchers have achieved ever-increasing photoreal-

ism from deep-learning techniques with the advent of

computer-generated imagery (CGI). CGI techniques

especially face synthesis, a.k.a deepfakes, pose an in-

creasing threat to manual and automatic face recogni-

tion systems. To overcome the risks posed by deep-

fakes, researchers are developing deep-fake detection

(DFD) methods. Recently, the digital AVATAR from

FaceMe (Fac, 2018) was deployed at Auckland Air-

port to answer bio-security questions (Auc, 2018).

The main effort in producing photo-real digital hu-

mans currently lies in face synthesis. The amount

of realism in face-synthesis is increasing over time.

The initial techniques for a deep-fake generation were

based on computer graphics posing it as a forward

face synthesis problem such as the one done by au-

thors in (NVi, 2011) where they produced a life-like

rendering of a human head requiring an expensive

setup. Hu et al. (Hu et al., 2017), and Yamaguchi

et al. (Yamaguchi et al., 2018) have approached face-

synthesis as a forward-technique using an expensive

setup called Light-Stage (Deb, 2000).

With the advent of deep learning, the requirement

for an expensive setup for face-synthesis has reduced

signiﬁcantly. Karras et al. (Karras et al., 2018) did one

of the initial works in this direction using Generative

Adversarial Networks (GANs). The quality of syn-

thesized images was further improved by Juefei-Xu et

al. (Juefei-Xu et al., 2018). Karras et al. (Karras et al.,

2019) also enhanced the quality of face-synthesis pro-

duced even more photo-real digital face. When we

say that face quality has improved during synthesis, it

needs to be pointed out that it includes hair and freck-

les of the skin. The results from Karras et al. (Karras

et al., 2019) are available on a public website (per, ).

However, recently the limitation of using an expen-

sive setup for acquiring Light Field for face-synthesis

was overcome in work by Sengupta et al. (Sengupta

et al., 2021) where it is obtained from a person watch-

ing videos on a regular computer.

DeepFakes (dee, 2019), one of the popular tech-

niques for replacing one person’s face with another,

leverages computer graphics & visualization tech-

niques and has been used for defaming persons. We

propose the use of DLDFD architecture based on deep

learning, and the contributions of our proposed ap-

proach are as follows:

• Our proposed approach is based on a recurrence-

free 2D convolution architecture DLDFD, which

involves augmented layers followed by Resnet-50

architecture, unlike the previous SOTA based on

3D convolution or recurrence-based 2D convolu-

tion.

568

Singh, J. and Ramachandra, R.

DLDFD: Recurrence Free 2D Convolution Approach for Deep Fake Detection.

DOI: 10.5220/0010880500003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages

568-574

ISBN: 978-989-758-555-5; ISSN: 2184-4321

(a) Real Face (b) Deep Fakes (c) Face2Face (d) Face Swap (e) Neural Textures

Figure 1: Low-Resolution Image Manipulation from Faceforensics++ dataset (R

ossler et al., 2018). Note the low-resolution

of real face makes the classiﬁcation challenging as distinction real v/s fakes becomes difﬁcult.

• We perform an extensive evaluation of Face-

Forensics++ low-resolution dataset, including

combined image-manipulation techniques, sin-

gle image-manipulation techniques, and cross-

evaluation of image-manipulation techniques. It

needs to be pointed out that there are few related

works for low-resolution evaluation of FaceForen-

sics++ compared to extensive literature for Deep-

Fakes in general.

• In terms of results, we achieve an accuracy of

96.07% compared to 85.14% of SOTA for com-

bined image manipulation. We achieve a 100%

accuracy for neural textures in single image ma-

nipulation techniques. Finally, we achieved an ac-

curacy of 94.28% for the unseen category of face

swap, much higher than SOTA for cross-image

manipulation techniques.

In the rest of the paper, we provide the literature re-

view of the critical papers in Deep Fakes Detection in

Section 2, followed by the proposed method in Sec-

tion 3. We describe experimental setup & results in

Section 4, and conclude the paper by providing con-

clusions & future work for Deep Fakes Detection in

Section 5.

2 RELATED WORK

In this section, we present the related work in ma-

nipulated face images. Digital Face Manipulation

can be grouped into expression swap (face reenact-

ment) and identity swap, as mentioned by Rossler et

al. (R

ossler et al., 2019). Identity Swap techniques

replace a genuine user’s face with another person, in-

cluding methods like FaceSwap (Kowalski, 2016) and

DeepFakes (dee, 2019). Expression Swap techniques

don’t change the person’s identity but change a real

user’s expressions by expressions obtained by another

person, including Face2Face (Thies et al., 2016) and

NeuralTextures (Thies et al., 2019). The manipula-

tions of Face2Face, & FaceSwap are based on com-

puter graphics techniques, whereas Neural Textures,

& DeepFakes are based on machine-learning.

We now review the essential works in Deep Fake

Detection (DFD). Li et al. (Li et al., 2018) did one

of the initial works in DFD where they proposed the

use of temporal pattern of eye blinking using a com-

bination of Convolutional Neural Network (CNN)

and a long-term recurrent neural network (CNN) on

a custom dataset achieving an area under the curve

(AUC) of 0.99. Li et al. (Li and Lyu, 2019) over-

came the limitation of training pairs of real and fake

samples for DFD by using face warping artifacts for

generating deepfakes from real images. They used

CNN for deepfake detection of public datasets of

DeepfakeTIMIT (Korshunov and Marcel, 2018), and

UADFV (Li et al., 2018), and achieved an AUC of

97.4% on UADFV, and for DeepfakeTIMIT an AUC

of 99.9% for Low Quality, & 93.2% for High Qual-

ity. Li et al. (Li et al., 2020) achieved SOTA re-

sults for the FaceForensics++ dataset based on the

observation that there is a blending boundary in the

forged image, which is absent in the real image. They

achieved an accuracy of 99% on single image manip-

ulation techniques and 97-98% on cross image ma-

nipulation techniques, but their approach requires a

mask in addition to the image during training. Jung

et al. (Jung et al., 2020) proposed the use of Gener-

ative Adversarial Networks (GAN) for detection of

signiﬁcant eye-blinking patterns in a video and have

an accuracy of 87.5% around a different type of deep-

fake videos. Sun et al. (Sun et al., 2021) proposed

the use of geometric features for Deep Fake Detection

where they ﬁrst calibrate the geometric features to

achieve a more precise location. This is followed by

using a two-stream Recurrent Neural Network (RNN)

to extract temporal features to achieve an AUC of

99.9% on the Faceforensics++ dataset. However, this

method provides a single evaluation and does not

cross-evaluate manipulation techniques. A summary

of challenges for deepfake detection was mentioned

by Lyu et al. (Lyu, 2020) where it is noted that deep-

fakes datasets have visual artifacts present in them,

DLDFD: Recurrence Free 2D Convolution Approach for Deep Fake Detection

569

and performance evaluation of deepfakes algorithms

is binary classiﬁcation & uses ﬁxed techniques, un-

like real-world deepfakes. The videos on social me-

dia platforms like Facebook and Instagram are usually

stripped of metadata and compressed for bandwidth

optimization, making the deepfakes classiﬁcation dif-

ﬁcult. A comprehensive survey about DeepFakes is

done by Tolosana et al. (Tolosana et al., 2020) where

they provide an overview of current SOTA methods

for deepfake detection, datasets, and open challenges

for the area. Mirsky et al. (Mirsky and Lee, 2021)

provided an overview of both creation and detection

algorithms for deepfakes.

We now speciﬁcally focus on classifying low-

resolution manipulated face images, which is a chal-

lenging area. In a more recent work, Sabir et al. (Sabir

et al., 2019) used recurrent neural network (RCN)

for DFD on FaceForensics++ dataset (R

ossler et al.,

2019), and achieve an accuracy of 96.9% on Deep

Fakes, 94.35% on Face2Face, & 96.3% on FaceSwap

on low-resolution videos from the dataset. It needs

to be pointed out that low-resolution image classiﬁ-

cation is challenging in general, and for manipulated

face images in particular (Wang and Dantcheva, 2020)

due to the high compression factor. The high com-

pression factor of low-resolution image-manipulation

techniques is shown in Figure 1. We now de-

scribe the related work for low-resolution image-

manipulation techniques, where the current SOTA is

by Wang et al. (Wang and Dantcheva, 2020) where

they performed manipulated video classiﬁcation on

FaceForensics++. They used 3d convolution-based

CNNs for their proposed approach which included

3D Resnet, & 3D Resnext (Hara et al., 2018), and

I3D (Carreira and Zisserman, 2017). Wang et al. eval-

uated base networks of 3D Resnet, 3D Resnext, and

I3D without modiﬁcation which is a limitation of their

technique.

Furthermore, 3D convolution is more memory in-

tensive to train compared to 2d convolution-based

deep-learning networks. Liu et al. (Liu et al., 2021)

reduced the computation budget of 3D convolution by

the use of Spatial rich model (SRM) features. How-

ever, they generally evaluated the Faceforensics++

dataset and were not speciﬁc to low quality.

3 PROPOSED METHOD

In this section, we present our proposed method

where the proposed architecture is chosen to give high

accuracy with minimal increase in computational cost

over the Resnet-50 architecture. This is achieved by

the use of augmented layers followed by Resnet-50

architecture. The design of augmented layers, which

includes three 3 × 3 convolutions, followed by one

1×1 convolution, is inspired from the inception mod-

ule of the Inception Network (Szegedy et al., 2015).

The augmented layers in the proposed network archi-

tecture are followed by Resnet-50 (He et al., 2016),

and it is chosen due to its high generalization capa-

bility as pointed out by He et al. (He et al., 2020).

The augmented layers are selected so that the com-

putational cost does not increase signiﬁcantly, and we

don’t use the complete inception module. Our pro-

posed approach consists of the different stages as in-

dicated in the block diagram in Figure 2.

3.1 Face Detection & Image

Normalization

We process the video dataset (R

ossler et al., 2019)

to extract frames, followed by face detection using

MTCNN (Zhang et al., 2016). The input face during

training is passed through the following transforma-

tions, which include random horizontal ﬂipping and

normalizing the image to mean, & variance of the Im-

agenet dataset (Krizhevsky et al., 2012). The main

idea for this step is to remove background as this can

increase the chances for the proposed deep network

to misclassify. The image normalization is performed

so that real and fakes have normalized data, resulting

in better classiﬁcation.

3.2 Proposed CNN Architecture

(DLDFD)

Our architecture consists of taking the input image

at a resolution of 224 × 224, which is followed by

three convolution layers (3 × 3 × 64), which are con-

catenated and passed through 1 × 1 Convolution (

1 × 1 × 192). This is followed by the use of Resnet-

50 (He et al., 2016) network architecture. The out-

put from Resnet-50 is passed through a fully con-

nected layer to generate labels. The main idea be-

hind the use of augmented layers at the top of the

Resnet-50 architecture is to allow the use of spar-

sity (Szegedy et al., 2015) in the proposed convolu-

tion neural network, which results in improved per-

formance with minor gain in the computational cost.

Three additional convolution layers are used to min-

imize the computational cost and achieve improved

performance. Further, the choice of 1 × 1 Convolu-

tion (Lin et al., 2013) results in dimensionality reduc-

tion and allows for patch-wise discrimination where

the latter is important for deepfakes classiﬁcation.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

570

Figure 2: Proposed Architecture for Manipulated Image Classiﬁcation.

3.3 Video Classiﬁcation

Once the network is trained, we extract the prediction

label from the fully connected layer for each video

frame. This is then followed by majority voting to

compute the video-level predictions. The loss func-

tion used during training is weighted binary cross-

entropy, and we use Adam Optimizer with a base

learning rate of 0.001, the momentum of 0.9, and 25

epochs. We choose weighted binary cross-entropy

as the loss function as it helps in the unbalanced

classiﬁcation as mentioned by Wang et al. (Wang

and Dantcheva, 2020). The Resnet-50 architecture

weights are initialized by the model trained on Ima-

genet from He et al. (He et al., 2016).

4 EXPERIMENTAL SETUP &

RESULTS

We ﬁrst describe the public dataset FaceForen-

sics++ (R

ossler et al., 2018) used for the compari-

son of our proposed method with SOTA. FaceForen-

sics++ dataset consists of the face manipulations in

video format and has 1000 real videos and four im-

age manipulations, each with 1000 videos. It consists

of image manipulations of DeepFakes, FaceSwap,

Face2Face, and Neural Textures in high-resolution

and low-resolution formats. We perform the exper-

iments on Faceforensics++ Dataset after extracting

frames from the videos, and we divide the dataset into

training, validation, and test sets. The training set

consists of 367228 pristine images, and in manipu-

lated images, we have 291789 from neural textures,

367009 deep fake, 292320 face swap, and 366631

face2face manipulations. The validation set consists

of 68857 pristine images, and in manipulated im-

ages, we have 54617 from neural textures, 68664 deep

fake, 54624 face swap, and 68854 face2face manipu-

lations. The testing set consists of 73768 pristine im-

ages, and in manipulated images, we have 59670 from

neural textures, 73766 deep fake, 59672 face swap,

and 73770 face2face manipulations. The dataset is

summarized in Table 1, both in the form of images

and videos. We perform different experiments on the

Faceforensics++ dataset on the lines of those devised

by Wang et al. (Wang and Dantcheva, 2020) where we

report True Classiﬁcations Rates (TCR) in them.

4.1 Combined Manipulation

Techniques

We present the results in Table 2 with combined ma-

nipulation techniques. The training and the testing are

performed as real v/s combined fakes which include

DeepFakes (DF), Neural Textures (NT), Face2Face

(FF), and FaceSwap (FS). We pose the classiﬁcation

as a two-class problem of real v/s combined fakes.

Our proposed method outperforms the current state-

of-the-art (SOTA) in this evaluation, as shown in Ta-

ble 2.

DLDFD: Recurrence Free 2D Convolution Approach for Deep Fake Detection

571

Table 1: Details of Faceforensics++ Dataset used.

Images

Phase Pristine NT DF FS F2F

Training 367228 291789 367009 292320 366631

Validation 68857 54617 68664 54624 68854

Testing 73768 59670 73766 59672 73770

Videos

Training 720 720 720 720 719

Validation 140 140 140 140 140

Testing 140 140 140 140 140

Table 2: Results on Low-Resolution Images from Faceforensics++ Dataset, with image-manipulations of (DeepFakes (DF),

Neural Textures (NT), Face2Face (FF), and FaceSwap (FS)).

Combined Evaluation of Low-Resolution Faceforensics++

Algorithm Train and Test TCR%

Resnet (Wang

and

Dantcheva,

2020)

FS,DF,F2F,NT 83.86

ResneXt (Wang

and

Dantcheva,

2020)

FS,DF,F2F,NT 85.14

Proposed

Method

FS,DF,F2F,NT 96.07

Single Evaluation of Low-Resolution Faceforensics++

Algorithm DF% F2F% FS% NT%

3D Resnet (Wang and Dantcheva, 2020) 91.81 89.6 88.75 73.5

3D Resnext (Wang and Dantcheva, 2020) 93.36 86.06 92.5 80.5

I3D (Wang and Dantcheva, 2020) 95.13 90.27 92.25 80.5

RCNN w Densenet (Sabir et al., 2019) 96.9 94.35 96.3

Proposed Method 94.64 81.4 94.28 100

Cross Evaluation of Low-Resolution Faceforensics++

Train Test 3D

Resnet (Wang

and

Dantcheva,

2020)

Resnext (Wang

and

Dantcheva,

2020)

I3D (Wang

and

Dantcheva,

2020)

Proposed

Method

FF, DF, F2F NT 64.29 68.57 66.79 49.2

FS, DF, NT F2F 74.29 70.71 68.93 40.0

FS, F2F, NT DF 75.36 75.00 72.50 69.28

DF, F2F, NT FS 59.64 57.14 55.71 94.28

4.2 Single Manipulation Techniques

We present the results in Table 2 with single manip-

ulation techniques. The training and the testing are

performed as real v/s individual fakes, which includes

DeepFakes (DF), Neural Textures (NT), Face2Face

(FF), and FaceSwap (FS). Our proposed method out-

performs the current state-of-the-art (SOTA) in this

evaluation for FS and NT. We get 100% classiﬁca-

tion accuracy on Neural Textures, attributed to our

proposed network identifying the neural renderings.

Further, our proposed network performed on par with

SOTA for deep fakes and face swap. Thus, our pro-

posed network can recognize changes in the image

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

572

when either the identity changes or a neural render-

ing is performed. However, only on face2face, our

proposed method performs slightly lower than SOTA,

with only expression change in this. This could be at-

tributed to the fact that there are local changes at the

image level during expression transfer.

4.3 Cross Manipulation Techniques

We present the results in Table 2 with cross manipula-

tion techniques. The training is performed as real v/s

cross fakes, which includes DeepFakes (DF), Neural

Textures (NT), Face2Face (FF), and FaceSwap (FS),

for, e.g., training includes real images, and fake im-

ages from FF, DF, & F2F, and testing is performed

on NT. Our proposed method outperforms the current

state-of-the-art (SOTA) in this evaluation when face

swap is used as identity change happens during this

testing. It performs on par with SOTA when deep

fakes are used during test time. However, for the

evaluation, when neural textures or face2face are used

during test time, our proposed network cannot gener-

alize that well. This can be attributed to our proposed

network’s performance when an identity change hap-

pens during the test time, which is the case for deep

fakes, and face swap test tasks.

4.4 Analysis of Results

We now analyze the results presented in previous sub-

sections. Our proposed method works well on the

combined image-manipulation and single-image ma-

nipulation techniques (Table 2). The case of com-

bined and single image manipulation is the seen class

scenario, and DLDFD performs well mainly due to

augmented layers. DLDFD (Table 2) achieves com-

parable accuracy with SOTA for the cross image-

manipulation technique result and is better than SOTA

for face-swap image-manipulation. This is the unseen

class case, and DLDFD generalizes well only when

identity changes are used during test time.

5 CONCLUSIONS &

FUTURE-WORK

In this paper, we proposed the use of DLDFD archi-

tecture based on 2d convolution with augmented lay-

ers to achieve better results for the evaluation of mul-

tiple image-manipulation techniques, two categories

for the assessment of single image-manipulation tech-

nique, and one category for the evaluation of cross

image-manipulation technique compared with SOTA

based on 3d convolution (Wang and Dantcheva,

2020). 2d convolution in DLDFD achieves better per-

formance at a lower computational cost.

We would extend the current technique to improve

cross image manipulation results speciﬁcally to gen-

eralize for expression change during test time in fu-

ture work. We would perform a more extensive com-

parison with the method by Sabir et al. (Sabir et al.,

2019) as they have provided only a few results with

single image manipulation techniques. The modiﬁ-

cation to improve cross image manipulation results

would involve redesigning the proposed network ar-

chitecture and using appropriate loss functions specif-

ically for improving generalization.

ACKNOWLEDGMENT

This work is carried out under the partial funding of

the Research Council of Norway (Grant No. IKT-

PLUSS 248030/O70).

REFERENCES

This person does not exist.

https://thispersondoesnotexist.com/. (Accessed

on 08/26/2021).

(2000). http://gl.ict.usc.edu/LightStages/.

(2011). http://ict.usc.edu/prototypes/digital-ira/.

(2018). Auckland Airport deploys avatar to answer biose-

curity questions. https://bit.ly/3ISXJ4w.

(2018). Face Me. https://www.faceme.com/.

(2019). Deepfakes code.

https://github.com/deepfakes/faceswap. (Accessed on

08/26/2021).

Carreira, J. and Zisserman, A. (2017). Quo vadis, action

recognition? a new model and the kinetics dataset.

In proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition, pages 6299–6308.

Hara, K., Kataoka, H., and Satoh, Y. (2018). Can spa-

tiotemporal 3d cnns retrace the history of 2d cnns and

imagenet? In Proceedings of the IEEE conference

on Computer Vision and Pattern Recognition, pages

6546–6555.

He, F., Liu, T., and Tao, D. (2020). Why resnet works?

residuals generalize. IEEE Transactions on Neural

Networks and Learning Systems, 31(12):5349–5362.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J.,

Sadeghi, I., Sun, C., Chen, Y.-C., and Li, H. (2017).

Avatar digitization from a single image for real-time

rendering. ACM Trans. Graph., 36(6):195:1–195:14.

DLDFD: Recurrence Free 2D Convolution Approach for Deep Fake Detection

573

Juefei-Xu, F., Dey, R., Bodetti, V. N., and Savvides, M.

(2018). RankGAN: A Maximum Margin Ranking

GAN for Generating Faces. In Proceedings of the

Asian Conference on Computer Vision (ACCV).

Jung, T., Kim, S., and Kim, K. (2020). Deepvision: Deep-

fakes detection using human eye blinking pattern.

IEEE Access, 8:83144–83154.

Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2018). Pro-

gressive growing of GANs for improved quality, sta-

bility, and variation. In International Conference on

Learning Representations.

Karras, T., Laine, S., and Aila, T. (2019). A style-

based generator architecture for generative adversar-

ial networks. In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition, pages

4401–4410.

Korshunov, P. and Marcel, S. (2018). Deepfakes: a new

threat to face recognition? assessment and detection.

CoRR, abs/1812.08685.

Kowalski, M. (2016). Faceswap code. (Accessed on

08/26/2021).

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. Advances in neural information processing

systems, 25:1097–1105.

Li, L., Bao, J., Zhang, T., Yang, H., Chen, D., Wen, F.,

and Guo, B. (2020). Face x-ray for more general face

forgery detection. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion, pages 5001–5010.

Li, Y., Chang, M.-C., and Lyu, S. (2018). In ictu oculi: Ex-

posing ai created fake videos by detecting eye blink-

ing. In 2018 IEEE International Workshop on Infor-

mation Forensics and Security (WIFS), pages 1–7.

Li, Y. and Lyu, S. (2019). Exposing deepfake videos by de-

tecting face warping artifacts. In IEEE Conference on

Computer Vision and Pattern Recognition Workshops

(CVPRW).

Lin, M., Chen, Q., and Yan, S. (2013). Network in network.

arXiv preprint arXiv:1312.4400.

Liu, J., Zhu, K., Lu, W., Luo, X., and Zhao, X. (2021). A

lightweight 3d convolutional neural network for deep-

fake detection. International Journal of Intelligent

Systems, 36(9):4990–5004.

Lyu, S. (2020). Deepfake detection: Current challenges and

next steps. pages 1–6.

Mirsky, Y. and Lee, W. (2021). The creation and detection

of deepfakes: A survey. ACM Comput. Surv., 54(1).

ossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies,

J., and Nießner, M. (2018). Faceforensics: A large-

scale video dataset for forgery detection in human

faces. CoRR, abs/1803.09179.

ossler, A., Cozzolino, D., Verdoliva, L., Riess, C.,

Thies, J., and Nießner, M. (2019). Faceforensics++:

Learning to detect manipulated facial images. arXiv

preprint arXiv:1901.08971.

Sabir, E., Cheng, J., Jaiswal, A., AbdAlmageed, W., Masi,

I., and Natarajan, P. (2019). Recurrent convolutional

strategies for face manipulation detection in videos.

CoRR, abs/1905.00582.

Sengupta, S., Curless, B., Kemelmacher-Shlizerman, I., and

Seitz, S. M. (2021). A light stage on every desk.

CoRR, abs/2105.08051.

Sun, Z., Han, Y., Hua, Z., Ruan, N., and Jia, W. (2021).

Improving the efﬁciency and robustness of deepfakes

detection through precise geometric features. In Pro-

ceedings of the IEEE/CVF Conference on Computer

Vision and Pattern Recognition (CVPR), pages 3609–

3618.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Computer Vision and Pattern Recognition (CVPR).

Thies, J., Zollh

ofer, M., and Nießner, M. (2019). De-

ferred neural rendering: Image synthesis using neu-

ral textures. ACM Transactions on Graphics (TOG),

38(4):1–12.

Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., and

Nießner, M. (2016). Face2face: Real-time face cap-

ture and reenactment of rgb videos. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 2387–2395.

Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A.,

and Ortega-Garcia, J. (2020). Deepfakes and beyond:

A survey of face manipulation and fake detection. In-

formation Fusion, 64:131–148.

Wang, Y. and Dantcheva, A. (2020). A video is worth more

than 1000 lies. comparing 3dcnn approaches for de-

tecting deepfakes. In 2020 15th IEEE International

Conference on Automatic Face and Gesture Recogni-

tion (FG 2020), pages 515–519. IEEE.

Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen,

W., Olszewski, K., Morishima, S., and Li, H. (2018).

High-ﬁdelity facial reﬂectance and geometry infer-

ence from an unconstrained image. ACM Trans.

Graph., 37(4):162:1–162:14.

Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint

face detection and alignment using multitask cascaded

convolutional networks. IEEE Signal Processing Let-

ters, 23(10):1499–1503.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

574