(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 13, No. 7, 2022

Fake Face Generator: Generating Fake Human Faces

using GAN
Md. Mahiuddin1, Md. Khaliluzzaman2, Md. Shahnur Azad Chowdhury3, Muhammed Nazmul Arefin4
Dept. of Computer Science and Engineering, International Islamic University Chittagong (IIUC) 1, 2, 4
Chattogram- 4318, Bangladesh1, 2, 4
Dept. of Computer Science and Engineering, Chittagong University of Engineering and Technology1
Chattogram- 4349, Bangladesh1
Dept. of Business Administration, International Islamic University Chittagong (IIUC), Chattogram- 4318, Bangladesh3

Abstract—As machine learning is growing rapidly, creating the fake images generated by the conventional GAN is
art and images by machine is one of the most trending topics limited. Wang et al. [2] in 2016 tried to simplify the overall
in current time. It has enormous applications in our current process of generative models which gave them more
day to day life. Various researchers have researched this topic realistic high resolution images as well as highly stable and
and they try to implement various ideas and most of them are robust learning procedures. Ghatas et al. [3] in 2020
based on CNN or other tools. The aim of our work is to proposed a method of building a complex modular pipeline
generate comparatively better real-life fake human faces with low using previously trained models to generate the kin image.
computational power and without any external image They tried to build such a model where it changes the way
classifier, rather than removing all the noise and maximizing the
of approaching GAN problems.
stabilization which was the main challenge of the previous
related works. For that, in this paper, we tried to implement Hamdi et al. [4] in 2019 proposed a new GAN that uses
our generative adversarial network with two fully connected the K-nearest neighbor for selective feature matching in a high
sequential models, one as a generator and another as a level space and they named their work as K-GAN. Tolosana
discriminator. Our generator is trainable which gets random et al. [5] in 2020 tried to recognize fake news, fake images,
data and tries to create fake human faces. On the other hand, fake media, and face manipulation using their method. They
our discriminator gets data from the CelebA dataset and tries provided a thorough review of digital manipulation techniques
to detect that the images generated by the generator are fake or
which were applied to facial content due to the huge number
real, and gives feedback to the generator. Based on the feedback
of possible detrimental applications.
the generator improves its model and tries to generate more
realistic images. Karras et al. [6] in 2019 proposed a method that re-designs
the generator architecture so that we can understand the various
Keywords—Generative adversarial network; fake human faces; aspects of the image synthesis process. Though GAN has seen
generator; discriminator; CelebA dataset rapid improvement in recent years, we still have a very
I. INTRODUCTION poor understanding of the properties of the latent spaces. They
tried to understand the inner workings of the latent spaces.
Fake images can be used by police to track down Their work starts learning from constant input and tries to
people who are involved in adultery. As we know nowadays adjust the “Style” of the image on each layer. Their method
child and underage pornography is alarming. To catch such was tested on Flickr-Faces-HQ and FFHQ datasets.
people we can use computer generated fake faces, that will
be more ethical than using real images [11]. As we can see Zhao et al. [7] in 2019 proposed an image translation
nowadays creating fake or edited images is very easy. To network by exploiting attributes with the generative adversarial
identify these kinds of images, we can build a model. For this network. It can remarkably contribute to the seven authenticity
model to be trained we can use the computer generated images of the generated face by supplementing the sketch image with
to identify the real and fake images. We use the fake computer the additional facial attribute feature. The generator and
generated images in visual art and the advertising industries. discriminator both use skip-connection to reduce the number of
These images have a huge role in computer games. In layers without affecting network performance. In the
computer and mobile games along with play station games, underlying feature extraction phase this network is different
these images can be used. As augmented reality and virtual from most attribute-embedded networks. They divided their
reality are growing rapidly, we can use these computer networks into two parts as sub-branch A and sub-branch B,
generated fake images. which takes a sketch image and attribute vector to extract low
level profile information and high level semantic features.
II. LITERATURE REVIEW
In recent years, Generative Adversarial Networks have
There are many fake image generation models created achieved extraordinary results for various applications in many
or updated by the researchers which generate pretty good fields especially in image generation because of their ability to
images. Zhang et al. [1] in 2020 discussed an algorithm that create sharp and realistic images. In this paper, Shirin Nasr
can improve the quality of generated images as the quality of Esfahani et al. [8] discussed five areas of image synthesis

160 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 7, 2022

based on different techniques of GAN. They are: Text to Image as we can see from the workflow diagram, we have used
Synthesis, Image to Image translation, Face Manipulation, 3D random noise as input data for the generator. We have also
Image Synthesis and Deep Master Prints. In this paper they used the Batch Normalization technique and the Reshape
have actually tried to focus on the applications of the above module from the keras to resize the data and normalize to train
techniques and discussed their merits and demerits and future the generator properly. In our generator we have used a total
applications. An Introduction to Image Synthesis with of five layers, one is an input layer with 128 nodes, starting
Generative Adversarial Nets is proposed in [10]. This model is with three more dense layers 512, 1024 and 2048 nodes
working well for the simple dataset, for the complex dataset the respectively with each layer and lastly the output layer. We
model does not work well. have used the Leaky ReLU activation function for all the
layers except the output layer; in the output layer we have
Though there are many fake image generation models, used the tanh activation function for the output layer. We
what we have tried to do in our work is that we have tried to have also used the loss function as the binary cross-entropy
generate the images using low computational power and for our generator model. Our generator will receive feedback
without any external image classifier. We have used two data from the discriminator and based on the received
neural networks in our model. One is Generator and the other feedback our generator will be updated.
is Discriminator. Generator creates fake human faces and
Discriminator detects the faces generated by the generator if On the other hand, our discriminator is also
this is fake or real. If the face generated by the generator is implemented with the sequential model. However, the main
detected as fake then the discriminator gives feedback to the difference between our generators is that after each epoch it
generator. And based on the feedback of the discriminator, does not learn, meaning it will not update or train after
the generator improves itself and tries to generate better each epoch. Our discriminator’s main purpose is to identify
quality faces. This process goes on until the generator the real and fake image and give the feedback to the
creates better quality fake human faces. From detecting fake generator, so that we will use the sequential model. In our
news, fake images to data misclassification, fake human face discriminator’s input data we have used both the generator
generation helps us to be used in all of these areas. And produced data and the real image from the CelebA dataset
generative adversarial network model is one of the best one and it takes both images and tries to identify the real and the
to create better real life fake human faces. fake image. Here, the model uses a total of four layers. As an
input layer we have taken the generated data and the real
Rest of the paper is organized as follows. The proposed image data and there are also two hidden dense layers along
method is explained in Section III. In Section IV, the with an output layer which produce binary values. Other than
experimental results are presented. The paper is concluded in the output layer we have used the Leaky ReLU activation
Section V. function in our model and as our output will be classification
III. PROPOSED METHOD type, so in the output layer we have used the sigmoid
function. For the loss function in our model we have used
A. Methodology the binary cross-entropy function.
In our model we have tried to implement generative B. Overall Workflow
adversarial networks in a very classic way with new and latest
tools and technologies. We have tried to implement a In Fig. 1 the diagram we have shown the overall
generative adversarial network to generate fake human faces workflow of our proposed model, and later in this paper we
from random noise with low computational power and have also explained how our generator, discriminator and the
without any external image classifier, rather than removing all fully connected sequential model work [12] [13].
the noise and maximizing the stabilization which was the main C. Generator
challenge of the previous related works. We know that for
implementing generative adversarial network we have to set We have proposed two sequential models, one as a
a generator whose work will be to generate fake faces from generator and another one as a discriminator to implement the
the noise and update its own model after receiving the generative adversarial model. We have used random noise as
feedback from the discriminator and we also have to input data in the generator model. The input data is
implement a proper discriminator to identify the real and fake actually a layer which has 100 nodes. Then we have used
four hidden layers where we have used Leaky Relu as our
image, so that it give the right feedback, means it says true to
activation function. The hidden layers have 128, 256, 512,
real image and false to fake image and based on its feedback
1024 nodes respectively. We have used the Tanh activation
the generator will change its model in such a way that
function in our last layer. We have also used Batch
eventually it will create almost real like images.
Normalization to normalize the data. At last, we have
That’s how we are trying to implement our generative reshaped the nodes into (48, 48, 1) size. The Adam optimizer
adversarial network to generate fake human faces. Our has been used as our optimizer. During the compilation of
generator and discriminator will use the keras’s classic the model, binary cross-entropy has been used as our loss
sequential modeling. function. The generator will produce some fake human
faces and it will improve itself by the feedback from the
We proposed to use two sequential models, one as a discriminator. The flow diagram of the generator is presented
generator and another one as a discriminator to implement our at Fig. 2.
generative adversarial model. For implementing the generator,

161 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 7, 2022

Fig. 3. Discriminator.

In our proposed model we have used two fully

connected sequential models, one as a generator and another
one as a discriminator. The generator receives a random
noise and tries to train it as a human face and then our
Fig. 1. Overall Workflow Diagram. discriminator receives data from the CelebA dataset and tries
to match the faces with the generator’s fake faces, and give
feedback to the generator, and based on the feedback our
generator learns and tries to create comparatively better fake
human faces after each training.
IV. EXPERIMENTAL RESULTS
In this section, we have presented a detailed overview of
the experiments of our presented model based on the CelebA
dataset. We have shown how our model worked during
different scenarios of the dataset. We have changed the size
of the data multiple times to see how the model works each
time. And we have got different results each time. We have
described how the results differ and improve over the time.
A. Dataset and Experimental Settings
We have used the CelebA dataset for our model. In
Fig. 2. Generator.
our CelebA dataset there are more than 200k data. At first,
we ran our model on this whole dataset but we used the size
D. Discriminator
of images as 48*48. We ran 100000 epochs on the model to
Discriminator is our second sequential model which we get a good result. Normally in generative adversarial models
have used in our generative adversarial model. At first we we have to run a lot of epochs to get a good result.
used the flatten operation to make the data one dimensional. Generative adversarial models generally generate new
After that there are two dense layers. The first hidden layer images according to the task given. So, it requires a lot of
has 48*48 nodes and the second hidden layer has half of the time to generate the new images. That’s why we ran the
first hidden layer, which means 1152 nodes. Leaky ReLU has model 100k epochs [14].
been used as the activation function in those two dense layers
having the learning rate as 0.2. In the output layer the sigmoid B. Experimental Tools and Environment
activation function has been used as the activation function. We have used different tools and environments for
We have used binary cross-entropy as the loss function and preprocessing the data and running the model so that we can
Adam as the optimizer during the compilation period. The have good results from our model.
discriminator will receive images from both the actual data
from the celebA dataset and fake data from the generator and C. Programming Language
it will give feedback to the generator if the images generated Tensorflow and Keras, which are the two libraries of
by the generator are fake or real. The generator will receive the Python programming language, have been used in our
feedback from the discriminator and improve itself. The flow model. Nowadays Python is the most popular programming
diagram of the discriminator is presented in Fig. 3. language for implementing Machine Learning and Deep
Learning models as it has many collections of packages

162 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 7, 2022

which are very helpful to implement the different Machine

Learning and Deep Learning models. Python is also the most
preferred language because it is very easy to learn and
implement. Python has many free and open source libraries
for Machine Learning and Deep Learning models.
Tensorflow is one of them. Keras is an open source software
library which supports a Python interface for artificial neural
networks. Keras plays the role as an interface for the
Tensorflow library.
D. Result
The four images shown in the Fig. 4 were generated by
the same model after 52000, 65000, 81000 and 97000 epochs
respectively. In the first image two faces could not be
generated properly. Rest of the faces of the first image were
generated but still there were a lot of noises which actually
prevented us from understanding the faces properly.
Compared to the first one, the second image is slightly a
little bit better but there are also one or two faces which
could not be generated properly and contained a lot of
noise. Almost all of the images were generated properly in the
third image containing some noise, though it was
comparatively better than the first two images. In the fourth
image, this image performed quite well. In the fourth image all
of the faces were generated and had a small noise. After
receiving the images from the CelebA database we first
flatten the images then we run the sequential model with the
Leaky ReLU. Then in the output layer we use the sigmoid
function to return the true or false type result value. Then
in the discriminator we use the binary cross entropy as the
loss function along with the Adam optimizer. This output
from the discriminator then goes to the generator as a
feedback and based on the feedback of the discriminator, the Fig. 4. Generated Data of 48*48 Size (CelebA Dataset).
generator tries to create new images. So, after the above
discussion we can say that there are still noises in the
generated faces at the end of the model too, though the results
of the latter part of the model performed comparatively better.
The four images shown in Fig. 5 were generated after
52000, 63500, 81500 and 99500 epochs respectively. After the
52000 epochs, two faces could not be generated properly,
which we can see from the first image above. Some faces are
not perfect though some faces are generated properly with
little noise. In the second image which was generated after
63500 epochs, all of the faces were almost properly generated
with some noise. After 81500 epochs and 99500 epochs, the
faces which were generated, that we can see from the
above third and fourth images, are much better. The
interesting fact is that if we observe the faces properly we can
see that almost all of the faces are different from one
another.
After the above discussion what we can observe is
that the faces became better as the time went on. And at the
last phase of the model the faces became more and more clear
and understandable. Another interesting fact is that as the
time went on all of the faces were generated by which we
can understand that the model was improving over time. And
we can also see the faces were different from one another Fig. 5. Generated Data of 100*100 Size (CelebA Dataset).
which is also a very positive output of this model.

163 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 7, 2022

E. Loss Function
Here, in the Fig. 6, we can see the loss function of the
discriminator. We ran our models at about 100000 epochs. If
we see the graph we can see most of the loss value between 0
to 1, which is pretty good value for the discriminator. And as
the discriminator is an un-trainable model we see the loss
value maintains the same ranges between the whole 100000
epochs. Now we can see some spike in the loss value of our
graph. If we try to separate the high spike by trying again
and again to visualize the diagram without the high spike of
the loss value in our (48 x 48) size images, we can then
easily see our loss value is somewhere between 0 and 1.4.
And if we calculate the high spike, we can see there are 771
high spikes between 100000 epochs which is around 0.77%
high spike and other 99.23% are the average, means 0 to
1.4 ranges. This is pretty good value for our model.

Fig. 8. Sample Output Images of the Baby Face Generation Model.

For sample output we have shown two figures. One is the

Fig. 6. Loss Values. Fig. 7 which is the generated output of our proposed model
and another is the Fig. 8 which is the generated output of the
F. Comparative Analysis baby face generation paper. At first glance we may think that
Zeliha Dogan [9] et al. proposed a model titled “Baby both of these papers are very different. However, if we think
Face Generation with Generative Neural Networks”, where clearly that both of those images are generated by a computer.
they tried to generate fake baby faces using comparatively And the objectives of those image creations are the same,
low computational power. We have chosen this paper for which means creating fake human faces. If we see the baby
comparison because both this paper and our work follow face generation model images, they clearly lack the
almost the same objectives. Both this paper and our model uniqueness of their faces. These images look like they were
mainly tried to generate fake faces using comparatively low created with some other faces. It’s more like editing rather
computation power. Below in this section we have tried to than creating or generating fake faces. On the other hand
compare our work with this paper, and show where our if we see our generated images we can clearly see that these
proposed method stands out and why. We have also tried to images are not like those. And surely we can’t figure out if
show our drawbacks and possible ways to overcome this these images are created by mixing some other faces. It’s
drawback. more like generating from the ground. Other than that we
can clearly see that our image has more sharpness in eyes,
nose and faces. But we have drawbacks in color and hair.
In the Table I, we can see that the different aspects of two
models. In our model we have used black and white data with
three different sizes of images 48*48, 100*100 and lastly
150*150. On the other hand, Baby face generation paper uses
RGB images with 200*200 sizes. They ran only 200 to 300
epochs whereas our model ran about 100000 epochs. In our
dataset we use 10k to 203k images on the other hand baby
faces use only 623 images.
Now we can clearly see that though the sizes of images are
smaller, we have used a lot of images with more epochs
without the color of the image which can be trained in
comparatively low computational power. Though we need
more computer power than the baby face generator paper,
using this extra computational power we can with proper
scaling of the images create better images with sharper faces.
Fig. 7. Sample Output Images of the Proposed Model.

164 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 7, 2022

TABLE I. COMPARISON TABLE

Number of Size of Dataset Generator

Model Image Type Image Size Da-taset Output
Epochs used Loss
Black & 48*48, 100*100, Overall better quality
Our Model 100000 CelebA 230K, 50K, 10K 0.0-10.0
white 150*150 image, sharp Face
Baby Face Generator Low quality image,
RGB 200*200 200-300 UTKFace 623 1.0-2.0
Model RGB image
discriminator to generate fake human faces. Though there
So after comparing with the baby face generation paper, are a lot of improvement areas in our proposed method, we
we can say that in our paper using some extra computation can say that using our proposed model we have successfully
power with proper image scaling we generated much better produced pretty good fake human faces comparatively using
human faces which are more realistic and sharper. But we very low computation power. In future, we can develop model
also have some drawbacks and our main drawback is that we to generate the images that can be used by police to track down
have used comparatively more computational power than the people who are involved in adultery. We can also develop
baby face generation paper and our resulting images are black models for generating fake computer generated images in
and white and our model needs comparatively more training visual art, for the advertising industries and computer games.
data.
REFERENCES
G. Discussion [1] Z. Zhang, X. Pan, X., S. Jiang, & P. Zhao, “High-quality face image
In this paper, we try to create fake human faces. Our generation based on generative adversarial networks,” Journal of
main focus was to create as much as a good image with low Visual Communication and Image Representation, Vol. 71, 2020.
computational powers without any external image classifier. [2] X. Wang, & A. Gupta, “Generative image modeling us- ing style and
structure adversarial networks,” In the European conference on
As we discussed above we used two sequential layers and computer vision, pp. 318-335, Springer, Cham, 2016.
a CelebA dataset to create fake human faces. [3] F. S.Ghatas, & , E. E. Hemayed, “Gankin: generating kin faces using
From our result and the loss function, we can see that, disentangled gan,” SN Applied Sciences, Vol. 2, No. 2, pp. 1-10, 2020.
we successfully created human faces from random noises [4] A.Hamdi, & B.Ghanem, “ IAN: Combining Generative Ad- versarial
Networks for Imaginative Face Generation,” arXiv preprint
using limited computational power. However, the images arXiv:1904.07916, 2019.
are not as good as the real life images, so there is lots of [5] R. Tolosana, R. Vera-Rodriguez, J. Fierrez, A. Morales, & J . Ortega-
scope in our paper to improve. However, we can see from Garcia, “Deepfakes and beyond: A survey of face manipulation and fake
the result that we have successfully created fake human detection,” Information Fusion, Vol. 64, pp. 131-148, 2020.
faces from the random noise data. This is one of our main [6] T. Karras, S. Laine, & T. Aila, “A style-based generator architecture
focuses. for generative adversarial networks,” In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pp. 4401-
Overall, using our low computational power we have tried 4410, 2019.
to generate real life fake human faces. For comparing this [7] J. Zhao, X. Xie, L. Wang, M. Cao, & M. Zhang, “Generating
model’s results and to see how it reacts with different sizes of photographic faces from the sketch guided by attribute using GAN,”
IEEE Access, Vol. 7, pp. 23844-23851, 2019.
images we have tested it with three different sizes. At first, we
[8] S. N. Esfahani, & S. Latifi, “Image generation with gans- based
tested it using 48 * 48 sizes. At that time, the model did not techniques: a survey,” AIRCC’s International Journal of Computer
give that much good result though it generated fake human Science and Information Technology, Vol. 11, No. 5, pp. 33-50, 2019.
faces having noises. And then, to improve the quality we [9] G. Ortaç, Z. Doğan, Z. Orman, & R . ŞAMLI, “ Baby Face
used 100*100 sizes. This time, the model performed quite Generation with Generative Adversarial Neural Networks: A Case
well, where it gave quite a good result. When we used Study,” Acta Infologica, Vol. 4, No.1, pp. 1-9, 2020.
100*100 sizes of the generated images were very clear and [10] H. Huang, P. S. Yu, & C. Wang, “An introduction to image syn- thesis
with generative adversarial nets,” arXiv preprint arXiv:1803.04469, 2018.
understandable. To get a better result, we then used 150*150
[11] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S.
images. This time the model gave comparatively better results Ozair, ... & Y. Bengio, “Generative adversarial nets,” Advances in
than the previous two models. Each time we ran our model neural information processing systems, Vol. 27, 2014.
at 100000 epochs. [12] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S . Ozair, ... & Y. Bengio, “Generative adversarial networks,”
V. CONCLUSION Communications of the ACM, Vol. 63, No. 11, pp. 139-144, 2020.
[13] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, &
In this work, a generative adversarial network model is A. A. Bharath, “Generative adversarial networks: An overview,” IEEE
proposed to generate fake human faces with low Signal Processing Magazine, Vol. 35, No. 1, pp. 53-65, 2018.
computational cost. The proposed generative adversarial [14] X. Liu, B. V. Kumar, Y. Ge, C. Yang, J. You, & P. Jia, “Normalised
network model generates comparatively better real life fake face image generation with perceptron generative adversarial
human faces with respect to the present state-of-the-art. We networks,” In 2018 IEEE 4th International Conference on Identity,
have also shown how we can only use two fully connected Security, and Behaviour Analysis (ISBA), pp. 1-8, IEEE, 2018.
sequential models, one as a generator and another one as a

165 | P a g e
www.ijacsa.thesai.org