Towards Fairness in Machine Learning: Balancing Racially Imbalanced
Datasets Through Data Augmentation and Generative AI
Anthonie Schaap
a
, Sofoklis Kitharidis
b
and Niki van Stein
c
Leiden Institute of Advanced Computer Science, Leiden University, Einsteinweg 55, 2333 CC Leiden, Netherlands
Keywords:
Fairness, Fair Machine Learning, Racial Bias, Inclusive AI, GANs, Dataset Balancing, Ethnicity
Classification.
Abstract:
Existing AI models trained on facial images are often heavily biased towards certain ethnic groups due to
training data containing unrealistic ethnicity splits. This study examines ethnic biases in facial recognition
AI models, resulting from skewed dataset representations. Various data augmentation and generative AI tech-
niques were evaluated to mitigate these biases, employing fairness metrics to measure improvements. Our
methodology included balancing training datasets with synthetic data generated through Generative Adver-
sarial Networks (GANs), targeting underrepresented ethnic groups. Experimental results indicate that these
interventions effectively reduce bias, enhancing the fairness of AI models across different ethnicities. This re-
search contributes practical approaches for adjusting dataset imbalances in AI systems, ultimately improving
the reliability and ethical deployment of facial recognition technologies.
1 INTRODUCTION
As artificial intelligence (AI) continues to evolve and
integrate into daily life, the need for transparent and
fair AI systems becomes increasingly critical. Cur-
rent AI systems can exhibit unwanted biases, often
due to biases present in the datasets they are trained
on (Srinivasan and Chander, 2021). These biases can
significantly impact real-world applications, where
human-like biases often arise due to imbalances in
ethnicity, age, or gender distribution within the train-
ing data. Biases can arise from uneven distribution of
attributes like age, ethnicity, and gender, as discussed
in (Wang et al., 2019). It is essential to tackle the
problem of model bias, which can have real-life im-
plications, such as in AI-driven hiring processes (Bo-
gen and Rieke, 2018) or risk assessments predicting
the likelihood of a defendant reoffending (Mehrabi
et al., 2022). In this paper, ethnicity bias in datasets
and AI models is investigated. To detect ethnicity
bias in datasets, an AI model is first trained to clas-
sify different ethnicities. This model requires miti-
gation of ethnicity bias to minimize its influence on
classification accuracy. Datasets for ethnicity clas-
a
https://orcid.org/0009-0007-4961-749X
b
https://orcid.org/0009-0005-8404-0724
c
https://orcid.org/0000-0002-0013-7969
sification can contain a lot of unwanted biases, and
are often unbalanced, as seen in Section 3.1. This re-
search aims to explore the presence of ethnicity bias
in datasets not specifically used for ethnicity classifi-
cation but for other purposes, such as age classifica-
tion and person re-identification. When heavy ethnic-
ity biases are present in a dataset but are unknown,
researchers might develop biased and racial AI mod-
els that may better classify certain ethnicities more
represented in the dataset. Historically many prob-
lems such as the Google Photos misdemeanor were
coupled with people of color, dehumanizing them and
mislabeling them. This research addresses the follow-
ing questions:
1. To what degree do existing face datasets and AI
systems show biases related to ethnicity?
2. How can fairness within facial datasets be im-
proved through diverse modifications to the train-
ing data?
3. Which method of adding facial images to the
training data best mitigates unwanted bias and im-
proves fairness?
Following the introduction of these key concepts
discussed in this paper, similar areas of research re-
lated to this topic are examined, found in Section
2. After a review of several state-of-the-art papers
on ethnicity bias and explainability, the methods pro-
584
Schaap, A., Kitharidis, S. and van Stein, N.
Towards Fairness in Machine Learning: Balancing Racially Imbalanced Datasets Through Data Augmentation and Generative AI.
DOI: 10.5220/0013002600003837
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Computational Intelligence (IJCCI 2024), pages 584-592
ISBN: 978-989-758-721-4; ISSN: 2184-3236
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

posed are outlined in Section 3, where the datasets
used in our experiments are also described. In Sec-
tion 4, the different experimental setups are described,
including the training of the ethnicity classification
model and the use of a generative AI model in com-
parison with several state-of-the-art approaches. The
results of these experiments, which compare our pro-
posed method to existing methods, are also presented
within this section. Following the presentation of
these results, a discussion of the findings is provided.
Lastly, conclusions are drawn and some limitations
and suggestions for future work are given in Section
6.
2 RELATED WORK
2.1 Bias in AI and Facial Recognition
Extensive research has highlighted the prevalence of
ethnic biases in AI systems. (Angwin et al., 2016)
found that algorithms used in the criminal justice
system disproportionately predicted more severe out-
comes for Black individuals, although (W Flores
et al., 2016) disputed these findings based on statis-
tical flaws.
In facial recognition, biases often arise from
dataset imbalances (Islam, 2023), where racial stereo-
types skew the representation of certain groups.
Techniques such as detecting segregation patterns in
datasets have been proposed to reduce disparities
among different ethnic groups (Benthall and Haynes,
2019). Additionally, (Alvi et al., 2019) introduced a
joint learning and unlearning algorithm that improved
fairness and accuracy by addressing biases related to
gender, age, and ethnicity in neural networks.
Advanced facial recognition frameworks such as
DeepFace (Serengil and Ozpinar, 2021), VGG-Face2
(Cao et al., 2018), and FaceNet (Schroff et al., 2015)
have been used for ethnicity classification tasks. Pre-
trained models like ResNet50 have also demonstrated
robust performance in classifying ethnicity and gen-
der attributes (Acien et al., 2019).
2.2 Methods for Mitigating Bias in AI
Mitigating bias in AI requires proactive measures to
identify and reduce biases in models (Howard and
Borenstein, 2018). (Raji et al., 2020) proposed a
framework to improve AI accountability by detecting
biases, while (Danks and London, 2017) emphasized
that model architecture can also harbor biases, neces-
sitating careful design and testing. Additionally, tech-
niques like synthetic faces can reveal unnoticed biases
related to appearance (Balakrishnan et al., 2020).
Generative Adversarial Networks (GANs) are a
promising tool for mitigating bias by generating sam-
ples representing underrepresented groups (Maluleke
et al., 2022). Advances like StyleGAN have improved
the generation of photorealistic images, allowing re-
searchers to manipulate key features such as pose and
expression (Karras et al., 2019). Building on this,
(Nitzan et al., 2020) introduced a method to separate
facial identity from other attributes, which we utilize
in our research to create diverse labeled data and im-
prove fairness in model training.
3 PROPOSED METHODOLOGY
Three methods are proposed and evaluated to miti-
gate bias by altering the training data during or before
model training. This section also outlines the datasets
critical for our methods.
3.1 Datasets
Experiments are conducted using various datasets tai-
lored for ethnicity classification. The first dataset,
UTKFace (Zhang et al., 2017), comprises over 20,000
images labeled with age, gender, and ethnicity. This
dataset is primarily used for training our ethnicity de-
tection model, featuring diverse labels. Labels in-
clude age (0 to 116 years), gender (0 for male, 1
for female), and race (0 for White, 1 for Black, 2
for Asian, 3 for Indian, 4 for Others). Particularly,
our focus is to address label imbalance, aiming for an
evenly distributed dataset across all racial categories.
The racial composition split is illustrated in Figure 1.
Figure 1: Different racial compositions in face datasets
from (K
¨
arkk
¨
ainen and Joo, 2019).
As depicted in Figure 1, the UTKFace dataset
does not have an evenly split on racial labels, but has a
far better split than other state-of-the-art face datasets,
excluding the FairFace dataset (K
¨
arkk
¨
ainen and Joo,
2019). Bias in training data, such as group imbal-
Towards Fairness in Machine Learning: Balancing Racially Imbalanced Datasets Through Data Augmentation and Generative AI
585

ances, typically leads to biased outcomes in AI mod-
els. Having a perfectly unbiased dataset is nearly im-
possible, but the goal is to mitigate the bias as much
as possible.
Additionally, Figure 1 highlights that, except
for FairFace, a disproportionately high number of
‘White’ images dominate most advanced datasets.
This imbalance can cause numerous issues; for in-
stance, when estimating the ages of individuals in a
dataset, the AI model is likely to perform better at
classifying images with the ‘White’ racial label due
to the greater volume of training data available for
this group, possibly resulting in poorer performance
for other racial groups.
To address this issue, experiments are also con-
ducted with the FairFace dataset (K
¨
arkk
¨
ainen and Joo,
2019) which provides an almost even split for the
racial labels White, Black, Latino, East Asian, South-
east Asian, Indian, and Middle Eastern. The FairFace
dataset contains 108.501 images, each labeled with
race, gender, and age. The racial compositions of the
UTKFace and FairFace datasets are presented in Ta-
ble 1. For clarity in presentation and analysis, the cat-
egories are grouped in the following manner in our ta-
bles: ‘White’, ‘Black’, ‘Asian’ (combining East Asian
and Southeast Asian), ‘Indian’, and ‘Others’ (encom-
passing Latino and Middle Eastern).
Table 1: Ethnicity split in the UTKFace and FairFace
datasets.
Ethnicity Count
UTKFace FairFace
White 8080 16527
Black 3621 12233
Asian 2759 23082
Indian 2146 12319
Others 1358 22583
3.2 Methodology
The following methods are proposed to sample or
generate additional data for underrepresented ethnic
groups.
Data Sampling Strategies
To use the datasets and train the model, some data
engineering is needed to match the data structures of
the different image datasets. Data preparation follows
the UTKFace dataset structure outlined in Section 3.1.
First, the FairFace dataset needs to be in the same
structure as the UTKFace dataset, where the image
labels are in the name of the images. After restruc-
turing, images are sorted into ethnicity-specific fold-
ers to facilitate targeted training on each class. This
splitting process is also done for the newly created
images using data augmentation and generative net-
works. For the generative images, the StyleGAN is
used to save the newly created images based on its
ethnicity. After the ethnicities of the created images
have been split, the name of every image is changed
to match the style of UTKFace. Names are updated
with a new age value and the appropriate ethnicity la-
bel. Timestamps are assigned with slight delays to
ensure unique filenames for each image. Each imple-
mentation can be used as a separate dataset or used
for additional training with specific ethnicity classes.
Data Augmentation
To increase both the size and quality of our datasets,
data augmentation is used to generate additional train-
ing data. Data Augmentation for image data is done
by adjusting the image slightly, such that the model
cannot recognize the image from its original state
and can use the image to improve its learning pro-
cess. Image adjustments include moving, rotating,
flipping, cropping, and shifting to diversify training
data. Color adjustments are also applied to prevent the
model from learning biases associated with specific
color schemes. Our aim is to keep augmented images
realistic, avoiding excessive vertical flips or rotations.
Additionally, the augmented images are saved to al-
low for manual quality control and future reuse. Table
2 displays the data augmentation methods used in this
research.
Table 2: Augmentation Methods Used to generate more re-
alistic images.
Aug. Method Value Description
Shear Range 0.05 Shear image by 5%
Zoom Range [1.0, 1.2] Zoom in up to 20%
Rotation 5 Rotate max. 5 deg.
Horizontal Flip [True,False] 50% flip chance
The augmentation values are randomly chosen
in range as shown in Table 2. The images are
not zoomed out, because that would create borders
around the image which are not part of the original
image. Moreover, image shifting is avoided to prevent
border creation and information loss. The augmented
images are labeled the same way the original images
are labeled, except with a new creation timestamp.
Generating Training Data Using StyleGAN
Another method to enhance our training dataset is by
generating new images using the attributes of exist-
ing ones. This is done using an implementation of
NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications
586

StyleGAN (Nitzan et al., 2020), where we use the in-
ference function to create new facial images based on
the identity and attributes of images from the training
set. For every combination of identity and attribute
image, a new image is created. For each classification
label, 2000 images are generated using 10 randomly
chosen attribute images and 200 identity images. This
expansion of the training set improves its stability and
quality, helping to prevent overfitting. When handling
small datasets, the use of StyleGAN to expand the
size of the training set can be very considerable, be-
cause the images created are different from the orig-
inal dataset. Generated images retain the ethnicity
label of the original identity image, ensuring consis-
tency in label integrity. This research also ensure that
the attribute image is from the same ethnicity label,
increasing the likelihood that the generated image be-
longs to that particular label class.
Among the various generative models evaluated,
StyleGAN was selected for its robustness and ease
of use. Other GANs were considered but not used
due to practical challenges such as unavailable model
weights, expired resource links, and substantial im-
plementation errors, which impeded reproducibil-
ity. StyleGAN’s capacity to manipulate attribute and
identity vectors allows for the generation of high-
quality, diverse images, facilitating the study’s focus
on enhancing dataset fairness.
The StyleGAN implementation (Nitzan et al.,
2020) utilizes a variety of pre-trained models devel-
oped using diverse datasets, such as FFHQ at resolu-
tions of 256x256 and 1024x1024 with 70,000 images
each, VGGFace2 with 3.31 million images, Celeb-
500K with 500,000 images, and 300W with 300 im-
ages. These pre-trained models enable efficient gen-
eration of new, diverse images, examples of which are
showcased in Figure 2.
3.3 Fairness Metrics for Performance
Evaluation
Various fairness metrics can be used to evaluate the
performance and fairness of the model. Models can
be evaluated based on their prediction accuracy for
each class, allowing for the determination of how
many true positives were correctly classified for each
class. However, in scenarios where classes are imbal-
anced, accuracy is not a good and informative met-
ric. It is important to also consider the false positives,
true negatives, and false negatives in the predictions.
This section examines different fairness metrics used
in this paper to evaluate fairness.
(a) White. (b) Black.
(c) Asian. (d) Indian.
Figure 2: Some generated images from the StyleGAN
model, using UTKFace images.
F1 Score
The F1 score, a harmonic mean of Precision and Re-
call, assesses classification performance, providing a
balance between the precision and recall for each la-
bel (Goutte and Gaussier, 2005):
Precision =
T P
T P + FP
, Recall =
T P
T P + FN
F1-score = 2 ×
Precision × Recall
Precision + Recall
The F1 score ranges from 0 to 1, with 1 indicating
perfect classification without error.
Equalized Odds
The Equalized Odds metric (Hardt et al., 2016) evalu-
ates model fairness by comparing True Positive Rates
(TPR) and False Positive Rates (FPR) across classes:
T PR =
T P
T P + FN
, FPR =
FP
FP + T N
The metric focuses on minimizing the maximum dif-
ference between the TPRs or FPRs across classes,
thereby enhancing fairness. It is given by:
Eq. Odds = max(|T PR1 − T PR2|, |FPR1 − FPR2|)
This research calculates Equalized Odds for all
class combinations, aiming to reduce the discrepan-
cies in recall and FPR, which could indicate bias. The
average of these differences is taken to represent the
model’s overall fairness in handling class disparities.
Towards Fairness in Machine Learning: Balancing Racially Imbalanced Datasets Through Data Augmentation and Generative AI
587

4 EXPERIMENTS
4.1 Ethnicity Detection Baseline Model
Experiments were conducted using a baseline model
for ethnicity detection, trained on the UTKFace
dataset. During 10 training epochs, the model
achieved a validation accuracy of about 91% and a
loss of 0.30.
The primary goal is not just high accuracy but
ensuring balanced accuracy across all ethnic groups.
This is critical for the model’s applicability across di-
verse datasets. A model may show high overall ac-
curacy but under-perform significantly on certain eth-
nicity labels, like ’Indian’, making it unfit for detect-
ing ethnicity bias. It is also essential that the model
avoids defaulting to a single ethnicity label when pre-
dictions are inaccurate.
Complete implementation details are available in
our code repository (Anonymous, 2024).
4.2 Effect of Additional Training
Methods
In this experiment, the equal FairFace dataset is used,
where each label has 4000 images. The goal is to cre-
ate a fair baseline, meaning that every label can be
classified with a similar accuracy. The similarity of
the classification accuracy is based on a difference of
2%, such that no label performs more than 2% better
than any other label. Ideally, the training set should be
ethnicity-balanced to achieve similar accuracy statis-
tics. In practice, this is often not the case, and ad-
ditional training is needed to balance the model. In
this experiment, we investigate the effect of adding
2000 images of a certain ethnicity label to the dataset,
such that the model can improve its learning process
on a particular label. Table 3 presents the performance
statistics of the baseline model on the UTKFace and
FairFace test sets.
Table 3: Baseline statistics for the Equal FairFace dataset,
tested on the FairFace and UTKFace test sets.
FairFace UTKFace
F1-score Acc F1-score Acc
White 0.59 52.0% 0.66 51.0%
Black 0.75 74.4% 0.81 76.2%
Asian 0.83 86.2% 0.83 85.9%
Indian 0.55 66.1% 0.68 66.2%
Others 0.52 47.8% 0.25 65.9%
As illustrated in 3, the baseline model trained on
the FairFace dataset is evaluated using two test sets.
This approach mitigates biases inherent in using a
test set derived from the same dataset. For UTKFace,
most of the images are labeled as ”White”, where the
same ethnicity split is used in the test set. This makes
the test accuracy far higher in comparison with us-
ing a test set with a balanced ethnicity split. It also
prevents the model from learning dataset-specific pat-
terns that could unreasonably influence test set accu-
racy. For the FairFace test set, the F1-score is the low-
est for the ”Indian” label, excluding the ”Others” la-
bel. To address this, 2000 data augmented ”Indian”
images were added to the original baseline training
data, and continued training. The results are presented
in Table 4.
Table 4: Additional training on ”Indian” label - FairFace
and UTKFace test set results.
FairFace UTKFace
F1-score Acc F1-score Acc
White 0.62 84.2% 0.86 84.2%
Black 0.72 58.3% 0.70 54.4%
Asian 0.80 90.2% 0.78 88.4%
Indian 0.52 64.3% 0.66 88.2%
Others 0.11 6.1% 0.05 2.7%
When adding data, it becomes evident that the
model stops trying to classify the ”Others” label, to
focus on the four more recognizable labels. The ”Oth-
ers” category consists of images that could be placed
in one of the four ethnicity labels, which makes the
”Others” category weak. Consequently, we exclude
the ”Others” category in subsequent experiments to
avoid its negative impact.
Equalized Odds on Real Data
A different method to evaluate the model involves us-
ing the Equalized Odds fairness metric from Section
3.3, aiming for the lowest possible Equalized Odds
value. A low Equalized Odds value signifies that True
Positive Rates (TPR) and False Positive Rates (FPR)
are similar across classes, enhancing fairness. After
calculating the Equalized Odds value between each
pair of classes, the average value is taken for each
class. Classes with the highest Equalized Odds values
receive additional training, as these indicate a need for
improved fairness. Figure 3 displays the outcomes
of the Equalized Odds experiment using real training
data.
Equalized Odds on Augmented Data
In this experiment, we change the additional train-
ing data to the augmented data from the UTKFace
dataset. The other hyperparameters like the number of
steps and the amount of additional images each step
NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications
588

Figure 3: Retraining the model using data derived from the
FairFace dataset. The additional training data is selected
based on the class with the highest Equalized Odds value,
aiming to enhance fairness across all ethnicity classes.
Figure 4: Retraining using augmented data derived from
UTKFace based on the highest Equalized Odds value. Af-
ter the initial training step, it is observed that the Indian
class exhibits the highest value, which means that this class
receives extra images based on data augmentation for the
next training step.
stays the same. Figure 4 presents the results of this
experiment.
Equalized Odds on GAN Data
In this experiment Generative Algorithms are used to
create new data for the training of the model. This
technique can be very helpful when dealing with very
small datasets, where a limited amount of images is
present. The Equalized Odds value for each ethnic-
ity class is calculated to demonstrate the effect of
adding additional training data to the baseline UTK-
Face training set. The results are displayed in Figure
5. Additional training is based on the highest Equal-
ized Odds value, aiming to balance the Equalized
Odds values over all classes. After several epochs,
Figure 5: Model retraining with GAN-generated data from
the UTKFace dataset, selected to improve fairness by tar-
geting classes with the highest Equalized Odds values.
it is observed that the Equalized Odds values of all
classes come fairly close to each other.
4.3 Balancing Fairness Using Different
Input Data
This section explores the impact of various ethnic
splits from different input datasets on model training.
There are differences between the original UTKFace
and FairFace models, with the possibility to add im-
ages using data augmentation, generative adversarial
networks, or simply using data from another dataset.
Baseline UTKFace
A method to balance fairness in machine learning
models is to adjust the input data the model uses for
training. This experiment examines the effect of alter-
ing the ethnicity split in the input data on the fairness
of the model. This is calculated using the Equalized
Odds fairness metric. As shown in Figure 6, the UTK-
Face dataset without the ‘Others’ category performs
reasonably well, particularly the ‘Indian’ class, which
exhibits the highest Equalized Odds value. An exam-
ination of the ethnicity split of the UTKFace dataset
in Table 1 reveals that the ‘Indian’ category contains
only 2146 images, whereas other classes have more
images. To address this imbalance, the next exper-
iment involves equalizing the image count for each
class to balance the dataset.
Equalizing Image Count in UTKFace and
FairFace
For the next experiment, images were added to the
Black, Asian, and Indian classes to match the num-
ber of images in the White class, which has the most
Towards Fairness in Machine Learning: Balancing Racially Imbalanced Datasets Through Data Augmentation and Generative AI
589

Figure 6: Equalized Odds values of UTKFace dataset with-
out the ”Others” category. It indicates that the ”Black” class
has the lowest average and max EQodds values, making it
the most ”fair” class in terms of Equalized Odds.
images with some 8063 different images. By using
an equalized image count for all classes, the poten-
tial for bias arising from an unbalanced ethnic split
in the input dataset is mitigated. This approach also
addresses the issue of the model potentially favoring
classes with more images. To augment the UTKFace
dataset and enable an equitable comparison with the
FairFace dataset, 4000 images of each class were se-
lected, totaling 16,000 images. This required adding
images to the Black, Asian, and Indian categories, as
detailed in Table 1, and removing excess images from
the White category by randomly selecting 4080 im-
ages for removal. After integrating randomly chosen
images from the FairFace dataset, each class was bal-
anced to exactly 4000 images. The test set, derived
from the original test set of the UTKFace dataset, was
also balanced by selecting 675 images of each class,
ensuring an equal number of test images for each pos-
sible classification. The results of this experiment are
displayed in Figure 7.
4.4 Improving Fairness Using
Additional Training by Adding
(n)Generated Images
This experiment examines the effect of adding images
from different datasets, augmented images or gener-
ated images using a GAN. This is done by adding
1000 images of the class with the lowest Equalized
Odds value, to achieve an as balanced as possible
fairness result using equalized odds. The experiment
is conducted over a single epoch during which the
adjusted UTKFace model is trained. After the first
epoch, the performance of the model is tested and the
fairness metrics are calculated. Based on the lowest
Figure 7: Equalized Odds values of the balanced UTKFace
dataset. The average values are averaged over 10 runs of
the model, with the maximum value found on the lighter
respectable color bars.
Figure 8: Equalized Odds values when performing different
additional training methods to the UTKFace dataset. Each
experiment is averaged over 10 runs. The area around the
lines is the confidence interval calculated by taking the stan-
dard deviation over the 10 runs.
Equalized Odds value, 1000 images from three pos-
sible separate datasets are added. The first dataset
comprised original FairFace images; the second fea-
tured UTKFace images augmented as described in
Section 3.2. The third dataset included images gen-
erated by StyleGAN from original UTKFace images.
All images from these three datasets are split over
each class, allowing for images from a single class to
be added to the original dataset. The experiment con-
sists of three separate runs, each its own dataset for
additional training. Each of the three experiments is
averaged over 10 runs, such that randomness is low-
ered. The outcomes are depicted in Figure 8.
NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications
590

Figure 9: Equalized Odds values calculated over 10 differ-
ent runs using different input training data sets. The UTK-
Face dataset has its original amount of images for each eth-
nicity. The other datasets have 4000 images for each class,
totalling 16k images per dataset. This balances out the
dataset but loses some of the original images correspond-
ing to the ”White” label.
4.5 Balancing Training Data by Adding
Images Using Various Techniques
Section 4.4 described adding data to the model based
on training performance. In this experiment, different
adaptations of the UTKFace dataset are used to bal-
ance the training data and improve the equalized odds
values from each ethnicity. By balancing the datasets,
every class has the same amount of labeled images,
such that each class gets an equal amount of train-
ing. The goal of this experiment is to create fairness
in the UTKFace dataset and to see which of the meth-
ods can create fairness. In practice, extra data is not
always available, so by using data augmentation or by
generating new images using StyleGAN, the process
of balancing datasets can still be improved. Figure 9
displays the results.
5 DISCUSSION
The experiments indicate that extending training with
new image data improves Equalized Odds across
classes, as shown in Figures 4 and 5. The StyleGAN-
generated data performed particularly well when orig-
inal images were limited. However, initial additional
training on specific classes (Indian and White, as
per Figure 3) led to worse Equalized Odds for other
classes. Comparable Equalized Odds values were
achieved for all classes when training on augmented
data (Figure 4), and notably improved for the Indian
and Black classes. Using an equalized image count
in UTKFace and FairFace datasets (Figure 7) demon-
strated similar Equalized Odds for all classes, albeit
”Indian” and ”White” showed consistent bias. Equal-
izing image count effectively curbed the model’s ten-
dency to favor labels with a higher image count.
Additional training using varying techniques
(Section 4.4) initially decreased Equalized Odds, but
these values increased after four epochs (Figure 9).
Yet, it failed to elevate the Indian class’s Equalized
Odds above other classes due to the model’s bias to-
wards larger image counts. Results suggest GAN im-
age data, FairFace, and augmented data perform simi-
larly. To understand the effect of GAN data, different
data will be added before training to achieve balanced
images.
Finally, balanced training data using various tech-
niques revealed the UTK+FairFace dataset as the
worst performer, possibly owing to the StyleGAN
model’s training on unbalanced data. Meanwhile,
data-augmented images performed similarly to the
UTKFace dataset, suggesting data augmentation as a
viable option for balancing and bias reduction. Fur-
ther experimentations with different fairness metrics
are needed to understand the impact of these tech-
niques better.
6 CONCLUSION
This paper investigates techniques to enhance fairness
in facial image datasets and models. Through our
literature review, we found models often use biased
image datasets, limiting their global usability. In re-
sponse, the study explores adding data during train-
ing, with a focus on the class exhibiting the poor-
est fairness. Our results show that additional training
and balanced training data effectively improve fair-
ness. As shown in Figure 9, adding StyleGAN gen-
erated images yields the worst fairness results, likely
due to the model’s inherent bias. Future work should
examine whether newer GAN implementations can
enhance fairness. Enhancing fairness will likely in-
crease the usage of these models in real applications
where decisions should not be based on bias. This
emphasizes the need for further exploration and im-
provements in fairness as AI continues to permeate
real-world applications.
REFERENCES
Acien, A., Morales, A., Vera-Rodriguez, R., Bartolome, I.,
and Fierrez, J. (2019). Measuring the gender and eth-
nicity bias in deep models for face recognition. In
Vera-Rodriguez, R., Fierrez, J., and Morales, A., edi-
Towards Fairness in Machine Learning: Balancing Racially Imbalanced Datasets Through Data Augmentation and Generative AI
591

tors, Progress in Pattern Recognition, Image Analysis,
Computer Vision, and Applications, pages 584–593,
Cham. Springer International Publishing.
Alvi, M., Zisserman, A., and Nell
˚
aker, C. (2019). Turning
a blind eye: Explicit removal of biases and variation
from deep neural network embeddings. In Leal-Taix
´
e,
L. and Roth, S., editors, Computer Vision – ECCV
2018 Workshops, pages 556–572, Cham. Springer In-
ternational Publishing.
Angwin, J., Larson, J., Mattu, S., and Kirchner,
L. (2016). Machine bias: There’s software
used across the country to predict future crim-
inals. and it’s biased against blacks. ProP-
ublica. https://www.propublica.org/article/
machine-bias-risk-assessments-in-criminal-sentencing.
Anonymous (2024). Towards Fairness In Machine Learning
- experiments, code and data. https://doi.org/10.5281/
zenodo.12672289.
Balakrishnan, G., Xiong, Y., Xia, W., and Perona, P. (2020).
Towards causal benchmarking of bias in face analy-
sis algorithms. In Vedaldi, A., Bischof, H., Brox, T.,
and Frahm, J.-M., editors, Computer Vision – ECCV
2020, pages 547–563, Cham. Springer International
Publishing.
Benthall, S. and Haynes, B. D. (2019). Racial categories in
machine learning. In Proceedings of the Conference
on Fairness, Accountability, and Transparency, FAT*
’19, page 289–298, New York, NY, USA. Association
for Computing Machinery.
Bogen, M. and Rieke, A. (2018). Help wanted: an exami-
nation of hiring algorithms, equity, and bias.
Cao, Q., Shen, L., Xie, W., Parkhi, O. M., and Zisserman,
A. (2018). Vggface2: A dataset for recognising faces
across pose and age.
Danks, D. and London, A. (2017). Algorithmic bias in au-
tonomous systems. pages 4691–4697.
Goutte, C. and Gaussier, E. (2005). A probabilistic inter-
pretation of precision, recall and f-score, with impli-
cation for evaluation. In Losada, D. E. and Fern
´
andez-
Luna, J. M., editors, Advances in Information Re-
trieval, pages 345–359. Springer Berlin Heidelberg.
Hardt, M., Price, E., Price, E., and Srebro, N. (2016). Equal-
ity of opportunity in supervised learning.
Howard, A. and Borenstein, J. (2018). The ugly truth about
ourselves and our robot creations: The problem of bias
and social inequity. Science and Engineering Ethics,
24.
Islam, A. U. (2023). Gender and Ethnicity Bias in Deep
Learning. PhD thesis.
K
¨
arkk
¨
ainen, K. and Joo, J. (2019). Fairface: Face attribute
dataset for balanced race, gender, and age. ArXiv,
abs/1908.04913.
Karras, T., Laine, S., and Aila, T. (2019). A style-based
generator architecture for generative adversarial net-
works.
Maluleke, V. H., Thakkar, N., Brooks, T., Weber, E., Dar-
rell, T., Efros, A. A., Kanazawa, A., and Guillory, D.
(2022). Studying bias in gans through the lens of race.
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., and
Galstyan, A. (2022). A survey on bias and fairness in
machine learning.
Nitzan, Y., Bermano, A., Li, Y., and Cohen-Or, D. (2020).
Face identity disentanglement via latent space map-
ping. ACM Transactions on Graphics (TOG), 39:1 –
14.
Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru,
T., Hutchinson, B., Smith-Loud, J., Theron, D., and
Barnes, P. (2020). Closing the ai accountability gap:
Defining an end-to-end framework for internal algo-
rithmic auditing.
Schroff, F., Kalenichenko, D., and Philbin, J. (2015).
Facenet: A unified embedding for face recognition
and clustering. In 2015 IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR). IEEE.
Serengil, S. I. and Ozpinar, A. (2021). Hyperextended light-
face: A facial attribute analysis framework. In 2021
International Conference on Engineering and Emerg-
ing Technologies (ICEET), pages 1–4. IEEE.
Srinivasan, R. and Chander, A. (2021). Biases in ai systems.
Commun. ACM, 64(8):44–49.
W Flores, A., Bechtel, K., and Lowenkamp, C. (2016).
False positives, false negatives, and false analyses:
A rejoinder to “machine bias: There’s software used
across the country to predict future criminals. and it’s
biased against blacks.”. Federal probation, 80.
Wang, T., Zhao, J., Yatskar, M., Chang, K.-W., and Or-
donez, V. (2019). Balanced datasets are not enough:
Estimating and mitigating gender bias in deep image
representations.
Zhang, Z., Song, Y., and Qi, H. (2017). Age progression
regression by conditional adversarial autoencoder.
NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications
592