Preprint Version

ICAACE 2025

Detection of AI Deepfake and Fraud in Online Payments Using GAN-Based

Models
________________________________________________________________________________________________________
Zong Ke, Shicheng Zhou,Yining Zhou∗ , Chia Hong Chang, Rong Zhang

* Corresponding author: Yining Zhou, [email protected]

The paper will be published and indexed by IEEE at 2025 8th International Conference on Advanced Algorithms and Control
Engineering (ICAACE 2025)

Jan. 2025
 Detection of AI Deepfake and Fraud in Online
Payments Using GAN-Based Models
Zong Ke1
National University of Singapore, Singapore 119077 Chia Hong Chang 4
[email protected] University of Colorado Boulder, Boulder, CO 80309, USA,
[email protected]
Shicheng Zhou2
University of Minnesota，425 13th Ave SE, Minneapolis, MN Rong Zhang 5
55414, USA, [email protected] University of California, Davis, CA 95616, USA,
[email protected]
Yining Zhou3&∗
*
Texas A&M University, College Station, TX 77840, USA, Corresponding author: [email protected]
[email protected]

Abstract—This study explores the use of Generative Adversarial in the field of image generation[4]. This foundational
Networks (GANs) to detect AI deepfakes and fraudulent activities in framework has since been refined by numerous advancements,
online payment systems. With the growing prevalence of deepfake such as DCGAN (Radford et al., 2016), CycleGAN (Zhu et al.,
technology, which can manipulate facial features in images and 2017), and StyleGAN (Karras et al., 2019)[5-7]. These
videos, the potential for fraud in online transactions has escalated. iterations have progressively enhanced the fidelity, stability,
Traditional security systems struggle to identify these sophisticated and versatility of generated images, pushing the boundaries of
forms of fraud. This research proposes a novel GAN-based model what synthetic media can achieve. Notably, StyleGAN's ability
that enhances online payment security by identifying subtle to generate high-resolution and photorealistic images has
manipulations in payment images. The model is trained on a dataset
established it as a cornerstone technology, especially in
consisting of real-world online payment images and deepfake images
generated using advanced GAN architectures, such as StyleGAN and
applications like AI-based facial manipulation and image
DeepFake. The results demonstrate that the proposed model can forgery detection.
accurately distinguish between legitimate transactions and deepfakes, In the realm of AI deepfake technologies, tools such as
achieving a high detection rate above 95%. This approach DeepFake (Chesney & Citron, 2019) have emerged as defining
significantly improves the robustness of payment systems against innovations[8]. Leveraging advanced deep learning models,
AI-driven fraud. The paper contributes to the growing field of digital DeepFake facilitates seamless facial swapping, producing
security, offering insights into the application of GANs for fraud
images and videos that are nearly indistinguishable from
detection in financial services.
authentic content[9-13]. This remarkable realism not only
Keywords- Payment Security, Image Recognition, Generative boosts the credibility of synthetic media but also expands its
Adversarial Networks, AI Deepfake, Fraudulent Activities use cases from entertainment and social media to malicious
domains[14-15]. Unfortunately, the same technology that
I. INTRODUCTION AND LITERATURE REVIEW drives creative innovation has been exploited for fraudulent
purposes, including online scams and identity theft, posing
With the rapid development of artificial intelligence,
significant security challenges[16].
especially generative adversarial networks (GANs), AI-driven
facial manipulation technologies, known as deepfakes, have II. OBJECT AND SUBJECT OF RESEARCH
become widely used in entertainment and social media.
However, their malicious use, particularly in online payments, The focus of this research is on the application of
presents a significant security threat[1]. Traditional payment Generative Adversarial Networks (GANs) in detecting
security systems struggle to detect these sophisticated risks. deepfake-related fraud in online payments. The study utilizes a
This paper proposes a GAN-based approach to identify dataset consisting of real-world payment images and
deepfake activities and associated fraud, creating a robust AI-generated deepfake images to train and evaluate the model.
security framework for online payments. The study focuses on The primary goal of this study is to develop a GAN-based
training GAN models to detect subtle facial manipulations, model capable of detecting AI-generated deepfake images in
enhancing payment system security by mitigating AI-driven online payment systems. This model aims to improve fraud
fraud risks[2-3]. This research contributes to payment security detection accuracy and contribute to the development of more
and demonstrates the potential of AI in addressing emerging robust payment security frameworks. The research also seeks
digital threats. to evaluate the effectiveness of GANs in differentiating
The introduction of generative adversarial networks (GANs) between legitimate and fraudulent payment images under
by Goodfellow et al. (2014) marked a revolutionary milestone various conditions.
 By bridging artistic creativity and technological precision,
these developments underscore both the potential and perils of
GAN-driven innovations in modern digital ecosystems.
III. METHODOLOGY
A. Data
The dataset utilized in this study comprises real-world
online payment images and AI-generated deepfake images. The
real-world images consist of 5,000 samples sourced from
publicly available payment platform datasets, including Google
Open Images, Kaggle, AI Benchmark Datasets, and illustrative
cases from Alipay's development documentation and demo Figure 1. The distribution characteristics of the real payment
materials as Figure 1. For the AI-generated deepfake images, images
an additional 5,000 samples were created using the most To ensure the dataset’s diversity and representativeness, it
advanced GAN models, including StyleGAN and DeepFake, as incorporates a wide range of facial images covering various age
table 1 below. groups, genders, and ethnicities. Furthermore, the dataset
Table 1 Descriptive Statistics includes both legitimate transaction scenarios and fraudulent
activities, providing a balanced foundation for evaluating the
proposed detection framework. This comprehensive dataset is
Category Real Payment Images Deepfake Images
an essential tool for training and testing the model’s capability
Number of to distinguish genuine transactions from deepfake-related fraud
5,000 5,000 in online payment systems.
Samples
Age Range 6 - 75 years 6 - 75 years
Gender Ratio near 1:1 near 1:1
Asian, Caucasian,
Ethnicity Same as real images
African, Hispanic
Resolution 1024x1024 pixels 1024x1024 pixels

Figure 2. Facial Feature Encoder-Decoder

The data processing section primarily involves
transforming data from scratch to structure and from disorder
to standardization as Figure 2. In this step, techniques such as
web scraping, regular expressions, normalization, and
standardization are used. The processed data is typically
represented as a tensor (a multi-dimensional array), and the
shape of the data varies depending on its type.
B. Model design
The core model employed in this paper is a generative
adversarial network (GAN), a deep learning framework
comprising a generator (G) and a discriminator (D). Through
adversarial training, GANs are capable of creating synthetic
data that closely mimics authentic data. In recent years, GANs
have achieved remarkable success in image generation and
forgery detection, with applications ranging from image
inpainting and style transfer to high-quality facial synthesis.
These advancements have further extended to fields such as
facial swapping (AI deepfake), making GANs highly relevant
for detecting fraudulent activities in online payments. In the
proposed model, the generator and discriminator fulfill
complementary roles as Figure 3:
 train the model, tuning the learning rate and batch size to
maintain a balanced development of generator and
discriminator. The model training was conducted over 10,000
iterations to ensure sufficient convergence and effectiveness in
detecting fraudulent behaviors in online payment scenarios.
IV. RESULTS
After training the model, we evaluated the performance of
generator and discriminator. The generator successfully
produced exceptionally realistic deepfake images, while the
discriminator achieved an accuracy rate of over 95% on the
validation set, effectively distinguishing between legitimate
payment images and fraudulent deepfakes.
Figure 3. GANs model flow chart At different stages of training, performance of the model
Generator (G): Takes a set of random noise inputs to produce was assessed with essential metrics, including accuracy,
synthetic deepfake images. Discriminator (D): Accepts input precision, and recall. The results are as following Figure 4:
images and determines whether they are genuine payment
images (non-fraudulent) or AI-generated deepfake images. By
iteratively training the generator to create more convincing
fraudulent images and the discriminator to more accurately
distinguish between real and fake images, the GAN framework
evolves into a highly effective system for detecting AI
deepfakes in payment scenarios. This adversarial training
process enables the model to identify subtle differences,
ultimately ensuring robust protection against deepfake-driven
fraud in online payment systems.
In generative adversarial networks (GANs), the loss function
serves as a crucial element in directing the training process. We
utilize the standard cross-entropy loss function:

𝐦𝐢𝐧 𝑽(𝐺, 𝐷) = 𝐸𝑥~𝑃𝑑𝑎𝑡𝑎(𝑥) [log(𝐷(𝑥))] + 𝐸𝑧~𝑃𝑧(𝑧) [log(1 − 𝐷(𝐺(𝑥)))] (1)

𝐺

Here, x represents real images sampled from the true data

distribution 𝑃𝑑𝑎𝑡𝑎 , and z denotes random noise input sampled
from a predefined noise distribution 𝑃𝑧 . The generator G maps
z to a synthetic image G(x).
For the discriminator D, the objective is to maximize the
likelihood of correctly identifying real images as genuine and
synthetic images as fake. The discriminator’s loss function is:

𝐦𝐚𝐱 𝐕(G, D) = Ex~Pdata(x) [log(D(x))] + Ez~Pz(z) [log(1 − D(G(x)))] (2)

𝐃

By adversarially optimizing these loss functions, the

generator enhances its ability to create realistic images.
Meanwhile, the discriminator becomes better at differentiating
real and fake images. This dynamic ensures a balanced
competition, ultimately leading to a robust GAN model capable
of effectively detecting AI-generated fraudulent activities in
online payment scenarios.
Throughout the training, the generator and discriminator are
optimized alternately over several epochs. Referring to the
study by Qin et al. (2018), which benchmarked and compared
various loss functions, it was observed that the choice of loss
function is not critically decisive, as no single function
outperforms the others in all scenarios. GANs are capable of
learning effectively under different conditions. Based on this
insight, we employed the commonly used Adam optimizer to
Figure 4. Model results
 These results demonstrate that the GAN-based model As Figure 5, the results validate that the GAN-based model
exhibits high accuracy and robustness in detecting AI deepfake is highly effective at detecting fraudulent deepfakes in online
and fraudulent activities in online payments. payment systems, offering significant improvements to the
security of digital transactions.
The performance of the proposed GAN-based detection
model was assessed using multiple evaluation metrics to V. CONCLUSIONS
capture its effectiveness in distinguishing between genuine and
fraudulent payment images. This paper proposes a novel approach to detecting AI
deepfake and fraudulent activities in online payments using
Table 2 summarizes the detailed results of the performance generative adversarial networks (GANs). Through
evaluation, demonstrating the model's effectiveness across experimental validation, the model demonstrated strong
several key metrics. performance in both detection accuracy and efficiency,
significantly enhancing the security of payment systems.
Metric Value
Despite achieving notable results in identifying AI
Accuracy 96.20%
deepfakes and fraud, certain challenges remain, such as
Precision 95.50% addressing fraudulent activities in complex scenarios and
Recall 96.80% improving the model’s capability generalize to various domains.
F1-Score 96.10% Future studies could explore the integration of multimodal data,
such as voice recognition or behavioral biometrics, to enhance
AUC 98.20% the GAN model’s ability to detect fraud in a wider range of
As table 2, Overall Accuracy: The model achieved an scenarios. Additionally, the performance of the model in
impressive 96.2% accuracy on the test dataset, demonstrating real-time payment systems and its adaptability across various
its strong capability to correctly classify both real and deepfake domains could be further investigated. A deeper exploration
payment images. This high level of accuracy reflects the into addressing edge cases and complex fraud patterns will also
model’s generalization ability and robustness across varied be necessary to refine the model’s generalizability.
inputs. Precision: The precision of the model, which measures
the accuracy of positive predictions, was found to be 95.5%. REFERENCES
This indicates that the model effectively minimizes false [1] Selvaraj, A., Selvaraj, A., & Venkatachalam, D. (2022). Generative
positive rates, meaning it rarely misidentifies legitimate Adversarial Networks (GANs) for Synthetic Financial Data Generation:
transactions as fraudulent. Recall: The recall metric, which Enhancing Risk Modeling and Fraud Detection in Banking and
Insurance. Journal of Artificial Intelligence Research, 2(1), 230-269.
captures the proportion of actual fraudulent cases correctly
[2] Charitou, C., Dragicevic, S., & Garcez, A. D. A. (2021). Synthetic data
identified, reached 96.8%. This high recall rate highlights the generation for fraud detection using gans. arXiv preprint
model's strength in detecting most of the fraudulent deepfake arXiv:2109.12546.
transactions, ensuring that the majority of fraud is caught. [3] Alo, S. O., Jamil, A. S., Hussein, M. J., Al-Dulaimi, M. K., Taha, S. W.,
F1-Score: The F1-score, representing the harmonic mean & Khlaponina, A. (2024, October). Automated Detection of
between precision and recall, stood at 96.1%. This balanced Cybersecurity Threats Using Generative Adversarial Networks (GANs).
result confirms the model’s capacity to maintain both high In 2024 36th Conference of Open Innovations Association (FRUCT) (pp.
566-577). IEEE.
precision and high recall, ensuring an effective fraud detection
mechanism. Receiver Operating Characteristic (ROC) Curve [4] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D.,
Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances
and Area Under the Curve (AUC): The ROC analysis yielded in neural information processing systems, 27.
an AUC of 0.982, which further underscores the model’s [5] Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., &
excellent performance in discriminating between legitimate and Chen, X. (2016). Improved techniques for training gans. Advances in
deepfake transactions. This value suggests that the model neural information processing systems, 29.
consistently performs well across different decision thresholds. [6] Chu, C., Zhmoginov, A., & Sandler, M. (2017). Cyclegan, a master of
steganography. arXiv preprint arXiv:1712.02950.
[7] Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., & Aila, T.
(2020). Analyzing and improving the image quality of stylegan. In
Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition (pp. 8110-8119).
[8] Chesney, R., & Citron, D. (2019). Deepfakes and the new disinformation
war: The coming age of post-truth geopolitics. Foreign Aff., 98, 147.
[9] Ke, Z., & Yin, Y. (2024). Tail Risk Alert Based on Conditional
Autoregressive VaR by Regression Quantiles and Machine Learning
Algorithms. arXiv preprint arXiv:2412.06193.
[10] Sharma, M., & Kaur, M. (2022). A review of Deepfake technology: an
emerging AI threat. Soft Computing for Security Applications:
Proceedings of ICSCS 2021, 605-619.
[11] Chadha, A., Kumar, V., Kashyap, S., & Gupta, M. (2021). Deepfake: an
overview. In Proceedings of second international conference on
computing, communications, and cyber-security: IC4S 2020 (pp.
Figure 5. Confusion matrix for the GAN-based fraud 557-566). Springer Singapore.
detection model [12] Whyte, C. (2020). Deepfake news: AI-enabled disinformation as a
multi-level public policy challenge. Journal of cyber policy, 5(2).
[13] Qin, Y., Mitra, N., Wonka, P. (2020). How Does Lipschitz [17] Sufian, A. (2024). AI-Generated Videos and Deepfakes: A Technical
Regularization Influence GAN Training? Computer Vision – ECCV Primer. Authorea Preprints.
2020. ECCV 2020. Lecture Notes in Computer Science, vol 12361. [18] Kalpokas, I., & Kalpokiene, J. (2022). I CAN Do It: From GANs and
Springer, Cham. https://doi.org/10.1007/978-3-030-58517-4_19 Deepfakes to Art. In Deepfakes: A Realistic Assessment of Potentials,
[14] Arshed, M. A., Mumtaz, S., Ibrahim, M., Dewi, C., Tanveer, M., & Risks, and Policy Regulation (pp. 55-63). Cham: Springer International
Ahmed, S. (2024). Multiclass AI-Generated Deepfake Face Detection Publishing.
Using Patch-Wise Deep Learning Model. Computers, 13(1), 31. [19] Ben Aissa, F., Hamdi, M., Zaied, M., & Mejdoub, M. (2024). An
[15] Giudice, O., Guarnera, L., & Battiato, S. (2021). Fighting deepfakes by overview of GAN-DeepFakes detection: proposal, improvement, and
detecting gan dct anomalies. Journal of Imaging, 7(8), 128. evaluation. Multimedia Tools and Applications, 83(11), 32343-32365.
[16] Singh, S., Sharma, R., & Smeaton, A. F. (2020). Using GANs to [20] Khoo, B., Phan, R. C. W. (2022). Deepfake attribution: On the source
synthesise minimum training data for deepfake generation. arXiv identification of artificially generated images. Wiley Interdisciplinary
preprint arXiv:2011.05421. Reviews: Data Mining and Knowledge Discovery, 12(3), e1438.