Heuristic Feedback for Generator Support in Generative Adversarial
Network
Dawid Połap
a
and Antoni Jaszcz
b
Faculty of Applied Mathematics, Silesian University of Technology, Kaszubska 23, 44-100 Gliwice, Poland
Keywords:
Generative Adversarial Network, Neural Network, Heuristic, Image Processing.
Abstract:
The possibilities of using generative adversarial networks (GANs) are enormous due to the possibility of
generating new data that can deceive the classifier. The zero-sum game between two networks is a solution
used on an increasingly large scale in today’s world. In this paper, we focus on expanding the model of
generative adversarial networks by introducing a block with a selected heuristic algorithm. The additional
block allows for creating a set of features extracted from the discriminator. The heuristic algorithm is based
on the analysis of feature maps and extracting the position of selected pixels. Then they are clustered into
averaged sets of features and used on created images by the generator. If the specified number of points within
any set of features is higher than the threshold value, then the generator performs classical training. Otherwise,
the loss function is subject to the penalty function. The proposed mechanism affects the operation of the GAN
through additional sample analysis concerning containing specific features. To analyze the solution and impact
of the proposed heuristic feedback, tests were performed based on known data sets.
1 INTRODUCTION
Recent years have brought new neural classifiers and
new methods of training including federated learn-
ing. The huge development is caused by the need for
fast and accurate analysis in the applications of these
methods (Yang et al., 2022; Salankar et al., 2023). An
example of what is the Internet of Things, Web 3.0,
Industry 4.0 (Zhang et al., 2022), etc. In addition,
neural networks are trained on huge data sets, which
contributes to high efficiency, as well as networks that
enable the generation of new data. Of course, the
basic application of such data may be augmentation,
however, the achievements of artificial intelligence
have allowed the creation of solutions that enable the
generation of data that can be mistaken for the works
of people. Examples of this are image generators and
different types of CNN-based models (Shahriar, 2022;
Artiemjew and Tadeja, 2022; Xu et al., 2023), etc.
The described possibilities are a consequence of
creating a zero-sum game where the players are two
neural networks. One trains to classify data as real
and fake, and the other generates new samples to de-
ceive the other network. Such contention is called
a
https://orcid.org/0000-0003-1972-5979
b
https://orcid.org/0000-0002-8997-0331
generative adversarial networks (GANs) (Cai et al.,
2022). The very wide application of such a solution
has contributed to increasing the capabilities of neural
networks in the industry. On the other hand, scientific
research on this type of network and their rivalry is
one of the main trends in improving their shortcom-
ings as well as increasing the accuracy or speed of
achieving high values of evaluation metrics. Aug-
mentation allows to the creation of artificial data to
increase the number of samples that train neural net-
works. In the case of dedicated applications, creating
a large amount of data in training databases may be
unattainable in a short time. Hence, generators are
used to streamline the base building process as well
as to enable faster achievement of higher efficiency.
Such a solution was presented in (Scarpiniti et al.,
2022), where the authors used GAN to generate au-
dio data presented in graphical form. For this pur-
pose, the focus was on spectral representation, which
enabled the correct further classification of the data.
A similar application was the use of augmentation to
obtain equinumerous training sets in the problem of
diagnosing bearing faults (Liu et al., 2022a). GAN
training involves a pre-prepared database for identi-
fying true and false images. Hence, in (Hung and
Gan, 2022) the modification of the generator architec-
ture with a double encoder and decoder was shown.
862
Połap, D. and Jaszcz, A.
Heuristic Feedback for Generator Support in Generative Adversarial Network.
DOI: 10.5220/0012405800003636
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Conference on Agents and Artificial Intelligence (ICAART 2024) - Volume 3, pages 862-869
ISBN: 978-989-758-680-4; ISSN: 2184-433X
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

The purpose of the modification was to increase the
amount of data in the training set, which in the end
allowed to obtain a higher accuracy of the classifier
by more than 10%. GAN can contribute to network
learning by generating synthetic images indicative of
segmentation. Such an example of use is shown in
(Ali and Cha, 2022). The authors proposed a pro-
prietary model of the GAN network and performed
tests based on images of damage to concrete elements.
Based on the presented research results, the authors
indicate that GAN-type networks can surpass other
neural networks in accuracy. The topic of data gener-
ation and their reliability is discussed in (Mozo et al.,
2022). The researchers presented the methodology of
using a heuristic algorithm to select the best generator
during GAN training.
An interesting solution is the use of GAN net-
works to generate super-resolution images (Jia et al.,
2022). To enable this, the generator architecture
in GAN has been refined through three processing
blocks: a convolution block, an upsampling block and
an attention fusion block. Another solution is to use
the GAN model for the classification or labeling of
samples. An example of such use is the development
of the architecture with the addition of additional dis-
criminators (Liu et al., 2022b). The described so-
lution operates on a proprietary middle-generation
module, which is introduced between the generator
and the combined discriminator. the conducted re-
search showed that it is an effective and universal so-
lution in terms of adaptability to data sets. Another
use of GAN is to generate data to attack intrusion de-
tection systems (Lin et al., 2022). The proposed so-
lution is based on the use of a generator to modify
the original data. The authors of the research, based
on the results, indicate the superiority of the proposal
over other methods due to the much greater general-
ization of attacks. Again, in (Mira et al., 2022), GAN
was used to increase the reality of video-to-speech
models. The model is based on video-to-speech con-
version, and then the result is evaluated by a critic.
This solution increases the credibility and realism of
the obtained measurements. Such solutions are also
analyzed through the safety and correctness of the re-
sults. An example of research in this area is the imple-
mentation of an evaluation mechanism using a clus-
tering algorithm(Venugopal et al., 2022). The authors
drew attention to the cosine similarity of the data as-
suming the preservation of statistical properties.
Based on the analysis of the literature, it can be
seen that GAN-type architectures are a very important
element used in generating various data. This is a so-
lution to increase the accuracy of the network during
training the classifier due to the creation of new train-
ing data. Based on these motivations, in this research
paper, we present the architecture of the GAN net-
work extended with a dedicated analysis of results us-
ing a heuristic algorithm. The analysis of the heuristic
algorithm enables better adaptability of the generated
images to increase the efficiency of the generator. In
addition, the use of heuristic feedback allows to speed
up the generator’s operation. The main contributions
of this research are:
• heuristic feedback mechanism for image analysis,
• extending GAN architecture by additional block
to support generator,
• proposed GAN evaluation by known datasets and
comparison with state-of-art.
2 PROPOSED METHODOLOGY
GAN architecture is based on two networks: discrim-
inator D and generator G. The discriminator will
be trained using a database X =

x
0
, x
1
, . . . , x
|X|−1

.
The result of the classification of a i-th sample will be
marked as D(x
i
). The generator will be trained by a
random given as latent vector z, so the created sample
will be denoted as G(z). Evaluation of the generated
sample by discriminator will be D (G(z)). Moreover,
let us denote Z as a set of generated vectors. E(·, ·)
as an error function between a given two values. The
loss function for the discriminator will be defined as:
L
D
= E (D (x) , 1) + E (D (G(z)), 0) , (1)
and for the generator, it will be:
L
G
= E (D (G(z)) , 1) . (2)
Applying a binary cross-entropy formula, the above
loss functions will be changed as:
L
D
= −
∑
x∈X,z∈Z
log(D(x)) + log (1 − D (G(z))) (3)
and for the generator, it will be:
L
G
= −
∑
z∈Z
log(D (G(z))) . (4)
2.1 Heuristic Feature Extractor
In GAN architecture, the training process is per-
formed separately for the discriminator and the gener-
ator. Hence, we can consider training a discriminator.
The idea is to train the network to classify the incom-
ing sample as true or false. In the case of real sam-
ples, the convolutional neural network will extract the
features of images occurring in a specific class. To
Heuristic Feedback for Generator Support in Generative Adversarial Network
863

analyze them, we propose the use of a heuristic algo-
rithm that will enable the detection of those features
that the classifier treats as significant. To access these
features, feature maps from the last convolution layer
will be extracted after the training process ends in the
current iterations. The obtained feature maps for a
given image will be presented as a set of n images
S = {s
1
, s
2
, . . . , s
n
}.
The application of the heuristic algorithm is based
on finding selected features. For this purpose, the al-
gorithm performs two stages of action: environment
recognition, i.e. pixel analysis to identify the back-
ground of the image, and then analysis of pixels of
the opposite color. If the background of the image
is black, the heuristic analyzes white pixels. How-
ever, it should be noted that individual pixels do not
indicate a specific feature. To enable the heuristic to
analyze the image and return specific patterns repre-
senting the feature, a two-stage displacement is pro-
posed. The first is the location of the most intensive
areas, followed by a search and neighborhood analy-
sis that indicates the area’s surroundings. The exact
operation of the heuristics will be described on the
example of the selected algorithm, which is the fox
algorithm (Połap and Wo
´
zniak, 2021). It is an algo-
rithm inspired by the behavior of foxes when hunting
prey. We assume that each heuristic algorithm has two
basic parameters, which are the number of individu-
als N and the number of iterations T . The number
of individuals determines the number of points (here
pixels) that will be analyzed in one iteration. Within
each iteration, the algorithm moves the points accord-
ing to two mechanisms, which are global and local
displacement. Assume that, the initial population of
foxes is generated at random, so there will be a set of
pixels’ coordinates marked as P = {p
1
, p
2
, . . . , p
N
},
where p
i
= (x
i
, y
i
) on image I, so the i-th pixel val-
ues is I(p
i
) and x
i
∈ ⟨0, w⟩, y
i
∈ ⟨0, h⟩ (w, h is width
and height of the image I). At the beginning of each
iteration, the best individuals p
best
in the entire popu-
lation are found - the selection depends on the given
fitness function F(·). Each pixel in t-th iteration is
placed according to a global movement, which allows
the evaluation of pixels much further away from the
current position. It is made using the following equa-
tion:
p
t
i
= p
t−1
i
+

α · sign

p
t
best
− p
t−1
i

, (5)
where α ∈

0, d(p
t−1
i
, p
t
best
)

is a random coefficients
in the given range, where d(·) is Euclidean metric.
Then, each fox p
i
makes a selection regarding fur-
ther movement according to a random parameter µ ∈
⟨0, 1⟩:
(
Move closer if µ > 0.75
Stay and disguise if µ ≤ 0.75
. (6)
If the choice is the movement, then the value of the
subject’s field of view is determined by:
r =




a
sin(φ
0
)
φ
0

if φ
0
̸= 0
rand(0, 1) if φ
0
= 0
, (7)
which allows the individual to be moved to a better
position in the immediate vicinity as:
(
x
t
i
=
⌈
ar · cos(φ
1
)
⌉
+ x
t−1
i
y
t
i
=
⌈
ar · sin(φ
1
)
⌉
+
⌈
ar · cos(φ
2
)
⌉
+ y
t−1
i
, (8)
where φ
1
, φ
2
∈ ⟨0, 2π⟩ and a ∈ ⟨0, 1⟩. In the situation
that a choice is made to stay (Eq. (6)), then the indi-
vidual does not perform a local displacement.
Figure 1: Visualization of the heuristic algorithm (from left
to right): the original image, individuals in the heuristic al-
gorithm located important pixels, returned a set of features
in the given image.
For feature extraction, we propose to analyze fea-
ture maps from the discriminator. The first iteration of
heuristic algorithms analyzes the environment which
is the image. This is to determine the color of the
features you are looking for. To this end, the first it-
eration of the heuristic is evaluated against a simple
fitness function that returns the average value of the
components of the RGB model, that is:
F
1
(I, p
i
) =
1
3
∑
k∈{R,G,B}
I
k
(p
i
). (9)
It should be also noted, that in the case of a
grayscale image, the values of all three components
are the same, hence we can assume that the above av-
erage is equal to the value of any of the components.
For this function, there is no preferred best value, so
the best individual in the population will be selected
at random. After the first iteration, each individual
is evaluated and the average values of all individuals
are determined. If this value is less than 128 (half of
the color range in the RGB model), the background
color will be black. However, when the average value
is higher than 128, the background will be white and
the searched features will be black.
Subsequent iterations can analyze the image by
looking for features that may be significant for the
classifier. For this purpose, individuals are evaluated
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
864

Figure 2: Visualization of the operation of the GAN architecture with the proposed heuristic algorithm.
using the second fitness function (the higher the value,
the more important the pixel):
F
2
(I, p
i
, ξ) =
ξ
∑
j=−ξ
ξ
∑
l=−ξ
F
3
(I, (x
i
, y
i
), (x
i
+ j, y
i
+ l)),
(10)
where F
3
(·) is an auxiliary function that evaluates a
given pixel by determining whether it is a pixel in-
dicating a difference in neighboring values. This is
important because the boundary points will be inter-
esting, i.e.:
F
3
(I, p
1
, p
2
) =
(
1 if I(p
1
) > 128 and I(p
2
) < 128,
0 otherwise.
(11)
The above equations search on a white background,
but in the opposite case, majority and minority signs
will be reversed.
When the algorithm reaches the iteration limit, the
obtained points are clustered to group them according
to a certain number of features. The algorithm per-
forms the same operation on all other feature maps
and then selects those features that are duplicated in
more than half of the maps. In discriminant training,
such features are extracted from samples. When all
analyses are completed, the set of features for each
sample is compared to combine features for the same
classes. It is done by clustering into κ classes (in the
case when this number is large, then more feature sets
reaming to analyze by generator). Within each class,
the sets are combined and clustered to the initial num-
Algorithm 1: GAN with heuristic support.
Input: Generator, Discriminator, T
GAN
training iterations, T heuristic
iterations, N heuristic population size,
γ matching value
1 i := 0;
2 while i < T
GAN
do
3 Train discriminator;
4 Get feature maps;
5 Use heuristics to finding locate features
by T iterations with N points;
6 Combine all features into one set;
7 Reduce the number of feature sets using
clustering algorithm;
8 Generate random sample;
9 Evaluate sample by discriminator and
heuristic matching algorithm;
10 if sample is not covered by one of the
heuristic class more than γ% points then
11 Apply penalty function;
12 end
13 Modify weights and filters in generator;
14 i+ = 1;
15 end
ber of points - this makes it possible to obtain aver-
aged values of features within a given class. A sim-
plified visualization of the operation of the heuristic
algorithm is shown in Fig. 1.
Heuristic Feedback for Generator Support in Generative Adversarial Network
865

Figure 3: Chart of accurate measurements of the used GAN
model depending on the used parameters of the heuristic
algorithm.
2.2 Heuristic Feedback Mechanism for
Generator
The sets of features prepared by heuristics during the
training of the discriminator are finally used during
the training of the generator. After the sample is cre-
ated, it is first checked by feature sets of heuristics.
The check consists of superimposing points from a
single set on the generated image. The pixel values
under these points are then checked. The check con-
sists of verifying the white color, i.e. if the pixel has
a value greater than 128, the pixel is passed. this op-
eration is performed for all matrices. If there is a sit-
uation in which more than γ% points from any set are
covered, that is, the sample contains features of one
of the classes.
According to the adopted training model of the
generator, the sample is evaluated by the discrimina-
tor and the value of the loss function is determined.
In the absence of feature assignment by comparing
the sets obtained by heuristics, the loss function is in-
creased by 10%, which is understood as the penalty
function. A simplified scheme of operation is pre-
sented in Alg. 1
3 EXPERIMENTS
To evaluate the proposed method and analyze the co-
efficients, two classical databases, MNIST (LeCun,
1998) and Fashion-MNIST (Xiao et al., 2017), were
selected. Databases contain images of size 28 × 28.
The DCGAN (Dewi et al., 2022) model implemented
in Keras was used as the GAN architecture. All
tests were performed on the Intel Core i7-8750H with
24GB RAM and NVIDIA GeForce GTX 1050 Ti.
3.1 Coefficients Analysis
In the beginning, the heuristic coefficients were an-
alyzed with the assumption of obtaining the best
possible results. For this purpose, we used only
the MNIST dataset with the following parameters:
three different iteration values T = {50, 100, 200} and
the number of individuals in the population as N =
{10, 50, 100, 200}. The first analysis was to find out
the best value for the mentioned heuristic parame-
ters. Therefore, a constant value of γ = 0.5 was cho-
sen. This γ value means that at least 50% of the lo-
calized features (in one of the ten classes - it was
set up to create only ten clustered classes) should be
covered on the generated samples in order not to ap-
ply the penalty function. The training iteration was
set as 1000. All tests were performed ten times and
all results were averaged. The obtained charts are
shown in Fig. 3. The obtained results indicate that
the greater the number of individuals and iterations,
the higher the accuracy. With only 50 iterations, the
accuracy of the GAN with the methodology used in-
creases. However, even with 200 individuals, the ac-
curacy reached below 50%. Doubling the iteration
allowed for achieving nearly 15% better results when
using 10 individuals. By increasing them to 200, the
efficiency reached 60%. However, the best results
were seen using 200 iterations, where the accuracy
increased rapidly and was able to exceed 65%. How-
ever, it should be noted that the use of minimum iter-
ation values or the number of individuals means quite
random values due to too small several analyzed im-
age areas. Especially through the use of clustering of
these features, which with high randomness may re-
sult in generating sets of random features. Hence, the
use of a large number of individuals and iterations is
an important element.
In the next part of the conducted experiments, we
used 100 individuals with 200 iterations to analyze the
effect of the gamma coefficient. For this purpose, the
generator was trained with different parameter values
of these parameters: γ = {0.1, 0.2, . . . , 0.9, 1}. The
value of the coefficient equal to 0 means that there
is no proposed mechanism for using the heuristic al-
gorithm. In Fig. 4 the relationship between the av-
erage accuracy of the GAN and the applied value of
the coverage coefficient of the set of heuristic features
and the generated images is presented. The applica-
tion of the proposed mechanism achieves results sim-
ilar to its absence in the case of a coefficient value
below 0.4. This is because this is a very small point
coverage value and its rapid exceeding will not cause
changes in the value of the generator loss function.
Hence, at values equal to or greater than 0.4, a dif-
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
866

Table 1: Comparison of the accuracy on GAN models with and without the proposed mechanism.
Accuracy Discriminator loss Generator loss
MNIST
DCGAN 0,6934 0,59 0,96
DCGAN with RFOA 0,7445 0,58 0,82
Fashion-MNIST
DCGAN 0,5234 0,63 1,03
DCGAN with RFOA 0,6639 0,52 0,83
Figure 4: Graph of the average accuracy values to the used
gamma parameter.
ference can be seen. The best results were obtained
using values γ = 0.6, where the accuracy with such
a small number of training (1000 iterations). Accu-
racy results were obtained higher by 5% compared to
the lack of the proposed method. Higher values in-
dicate frequent use of the penalty function, which in
turn manifests itself in worse and worse accuracy.
3.2 Heuristic Analysis
The next stage of evaluating the proposals was the
use of other heuristic algorithms and the indication
of differences between them. Three selected heuris-
tics were implemented, such as RFOA (presented in
this paper), artificial rabbits optimization algorithm
(AROA) (Wang et al., 2022) and white shark opti-
mizer (WSO) (Braik et al., 2022). All algorithms
were identically modified to analyze the images by
adding the ceiling function to the equations and us-
ing the presented fitness functions. A variant of one
hundred individuals in each population and different
numbers of iterations were used. Each of the tests
for the selected algorithm was performed ten times,
and the results were averaged, as shown in Fig.5. All
selected algorithms obtained similar results, differing
from each other at the level of 0.2%. It should be
noted that AROA is the only one to have more jumpy
accuracies with the increasing number of iterations.
For the other two algorithms, the accuracy increases
almost linearly. It is worth noting that each heuristic
Figure 5: Dependence of the number of iterations for 100
individuals in selected heuristic algorithms with γ = 0.6 to
accuracy.
algorithm works on almost identical operations with
the difference related to the modeled equations. Con-
sidering that the same coefficient values were used in
the tested algorithms, the difference can be stated that
with smaller numbers of iterations, RFOA can locate
image features faster based on the declared fitness
functions.
3.3 GAN Analysis
The best parameters in previous tests turned out to be
RFOA with 200 iterations, 200 individuals and the
parameter γ = 0.6. To reduce the number of opera-
tions, the number of individuals was reduced to 100,
but the number of GAN training iterations was in-
creased threefold. Using these values, we train the
GAN model with and without the proposed heuris-
tic mechanism for two datasets: MNIST and Fashion-
MNIST. The obtained results are shown in Tab. 1.
In the case of the MNIST dataset, the accuracy for
the original GAN model was 69,34% with a generator
loss of 0,96. In the case of using the proposed mecha-
nism, the accuracy obtained was higher by more than
5%, as it reached 74,45%. A more important factor
is the generator loss, which is significantly lower -
0.82. Compared to the original classifier, this is a dif-
ference of 0.14. Again, for the second database, the
results turned out to be much weaker due to the analy-
sis of more complex objects in terms of features. The
Heuristic Feedback for Generator Support in Generative Adversarial Network
867

original GAN model achieved an accuracy of 52.34%
and the proposed algorithm 66.39%. This is a much
higher result, which is visible in the values of the loss
function. For the discriminator, there is a difference
of 0.11 (while for the previous database, it was 0.01),
and for the generator, it was 0.2. Based on the ob-
tained results, it can be concluded that the proposed
mechanism of image analysis in terms of having cer-
tain features is an effective solution that can signifi-
cantly affect the operation of GAN models.
4 CONCLUSIONS
In this paper, we presented a modification of the GAN
learning model. We extended the operations by build-
ing a set of essential features based on the images re-
turned by the discriminator by the heuristic algorithm.
The set is used by the generator to check whether the
generated image has essential features for the discrim-
inator. If not, the loss function is subject to the value
of the penalty mechanism. Based on the test results, it
was noticed that the proposed methodology allows for
a significant increase in the accuracy of the generator.
Especially when the database was much more diverse.
It was noticed that the modification of the value of the
loss function affects the training of the network due to
the learning algorithm, which is ADAM. In addition,
the implemented heuristic algorithm can achieve good
results with lower parameter values. The presented
method is important for enabling more efficient train-
ing of the generator in GAN models.
REFERENCES
Ali, R. and Cha, Y.-J. (2022). Attention-based generative
adversarial network with internal damage segmenta-
tion using thermography. Automation in Construction,
141:104412.
Artiemjew, P. and Tadeja, S. K. (2022). Using convnet for
classification task in parallel coordinates visualization
of topologically arranged attribute values. In ICAART
(3), pages 167–171.
Braik, M., Hammouri, A., Atwan, J., Al-Betar, M. A., and
Awadallah, M. A. (2022). White shark optimizer: A
novel bio-inspired meta-heuristic algorithm for global
optimization problems. Knowledge-Based Systems,
243:108457.
Cai, J., Li, C., Tao, X., and Tai, Y.-W. (2022). Image multi-
inpainting via progressive generative adversarial net-
works. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition, pages
978–987.
Dewi, C., Chen, R.-C., Liu, Y.-T., and Tai, S.-K. (2022).
Synthetic data generation using dcgan for improved
traffic sign recognition. Neural Computing and Appli-
cations, 34(24):21465–21480.
Hung, S.-K. and Gan, J. Q. (2022). Boosting facial emotion
recognition by using gans to augment small facial ex-
pression dataset. In 2022 International Joint Confer-
ence on Neural Networks (IJCNN), pages 1–8. IEEE.
Jia, S., Wang, Z., Li, Q., Jia, X., and Xu, M. (2022). Multi-
attention generative adversarial network for remote
sensing image super resolution. IEEE Transactions
on Geoscience and Remote Sensing.
LeCun, Y. (1998). The mnist database of handwritten digits.
http://yann. lecun. com/exdb/mnist/.
Lin, Z., Shi, Y., and Xue, Z. (2022). Idsgan: Genera-
tive adversarial networks for attack generation against
intrusion detection. In Pacific-Asia Conference on
Knowledge Discovery and Data Mining, pages 79–91.
Springer.
Liu, Y., Jiang, H., Liu, C., Yang, W., and Sun, W.
(2022a). Data-augmented wavelet capsule generative
adversarial network for rolling bearing fault diagnosis.
Knowledge-Based Systems, 252:109439.
Liu, Y., Wei, X., Lu, Y., Zhao, C., and Qiao, X. (2022b).
Source free domain adaptation via combined discrim-
inative gan model for image classification. In 2022
International Joint Conference on Neural Networks
(IJCNN), pages 1–8. IEEE.
Mira, R., Vougioukas, K., Ma, P., Petridis, S., Schuller,
B. W., and Pantic, M. (2022). End-to-end video-
to-speech synthesis using generative adversarial net-
works. IEEE Transactions on Cybernetics.
Mozo, A., Gonz
´
alez-Prieto,
´
A., Pastor, A., G
´
omez-Canaval,
S., and Talavera, E. (2022). Synthetic flow-based
cryptomining attack generation through generative ad-
versarial networks. Scientific reports, 12(1):1–27.
Połap, D. and Wo
´
zniak, M. (2021). Red fox optimiza-
tion algorithm. Expert Systems with Applications,
166:114107.
Salankar, N., Koundal, D., Chakraborty, C., and Garg, L.
(2023). Automated attention deficit classification sys-
tem from multimodal physiological signals. Multime-
dia Tools and Applications, 82(4):4897–4912.
Scarpiniti, M., Mauri, C., Comminiello, D., Uncini, A., and
Lee, Y.-C. (2022). Coval-sgan: A complex-valued
spectral gan architecture for the effective audio data
augmentation in construction sites. In 2022 Interna-
tional Joint Conference on Neural Networks (IJCNN),
pages 1–8. IEEE.
Shahriar, S. (2022). Gan computers generate arts? a sur-
vey on visual arts, music, and literary text generation
using generative adversarial network. Displays, page
102237.
Venugopal, R., Shafqat, N., Venugopal, I., Tillbury, B.
M. J., Stafford, H. D., and Bourazeri, A. (2022). Pri-
vacy preserving generative adversarial networks to
model electronic health records. Neural Networks,
153:339–348.
Wang, L., Cao, Q., Zhang, Z., Mirjalili, S., and Zhao,
W. (2022). Artificial rabbits optimization: A new
bio-inspired meta-heuristic algorithm for solving en-
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
868

gineering optimization problems. Engineering Appli-
cations of Artificial Intelligence, 114:105082.
Xiao, H., Rasul, K., and Vollgraf, R. (2017). Fashion-
mnist: a novel image dataset for benchmark-
ing machine learning algorithms. arXiv preprint
arXiv:1708.07747.
Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C.,
Gan, W., Yan, Y., Jin, X., Yang, X., et al. (2023).
Actformer: A gan-based transformer towards general
action-conditioned 3d human motion generation. In
Proceedings of the IEEE/CVF International Confer-
ence on Computer Vision, pages 2228–2238.
Yang, W., Xiang, W., Yang, Y., and Cheng, P. (2022).
Optimizing federated learning with deep reinforce-
ment learning for digital twin empowered industrial
iot. IEEE Transactions on Industrial Informatics,
19(2):1884–1893.
Zhang, T., Gao, L., He, C., Zhang, M., Krishnamachari,
B., and Avestimehr, A. S. (2022). Federated learning
for the internet of things: Applications, challenges,
and opportunities. IEEE Internet of Things Magazine,
5(1):24–29.
Heuristic Feedback for Generator Support in Generative Adversarial Network
869