Improved Accuracy in Deepfake Detection Using GAN and

Fisherface Algorithm

M. Udhaya Kumar, B. Latha, B. Vinoth Kumar, R. Srinithi, B. Elamathi and C. Soundharya

Department of Electronics and Communication Engineering, K S R College of Engineering, Tiruchengode – 637215,

Namakkal, Tamil Nadu, India

Keywords: CNN, Precision, Feature Extraction, Dimensionationality Reduction, Facial Recognition, Fisherface, GAN,

Deepfake.

Abstract: Aim: The goal of this paper is to design a robust deepfake detection method by integrating GAN, and the

Fisherface algorithm to increase accuracy and precision in the detection of fabricated media. The performance

of this model is compared with traditional CNN and LSTM-based models. This research is into two groups.

Group 1 is a CNN-LSTM with 950 samples in order to capture spatial and temporal features. Group 2 utilizes

GANs in synthetic data augmentation with Fisherface feature extraction and SVM classification with 1030

samples. Results: The hybrid GAN-Fisherface-SVM model results in significantly higher detection accuracy

compared to traditional models. The hybrid model shows a significant gain of detection, which stands out at

about 5-10%, by measurement matrices such as accuracy, error rate and response time with a significance

value below 0.05. Conclusion: Overall, the developed approach that combines data augmentation using a

GAN method along with a Fisherface algorithm performs a dramatic level of recognition towards deep fake

as compared with methods used.

1 INTRODUCTION

Driving its rise, deepfake technology uses

increasingly sophisticated machine learning

algorithms to produce hyper-realistic content, creating

other world concerns for cybersecurity and risk

management of both misinformation and news

authenticity (Korshunov, et.al.,2018) . These synthetic

videos, images and audio manipulations can

realistically transform reality, and it is becoming

increasingly difficult to detect in areas such as

politics, entertainment and law enforcement. Existing

techniques rely on facial features, sound, and motion,

which are susceptible to manipulation through

generation adversarial networks (Mirsky, et.al.,2021).

The GANs have been widely used in both generating

and detecting deepfakes, generating realistic synthetic

data to facilitate the accuracy of discrimination.

Fisherface algorithm is well known for being

computationally light in terms of facial identification,

which is in turn a critical aspect of deepfake detection

(Belhumeur, et.al.,1997). This paper proposes a

hybrid deepfake detection framework by exercising

GAN based data augmentation, Fisherface features

extraction, and SVM classification. The proposed

architecture involves a combination of GAN based

realistic data generation and Fisherface, the feature-

based recognition approach that improves the ego-net

detection accuracy compared to original samples of

deepfake images Smith, J., & Doe, A. (2023).

According to the evaluation results, despite

performing thorough research, we have significantly

reduced the accuracy, error rate and response time of

the detection compared to traditional methods.

Training on data collected until October 2023

empowers detection in progressively adversarial

settings, boosting the robustness of security aspects

such as forensic and media validation. Future work

may focus on further advanced optimizations for

increased robustness in deepfake detection.

2 RELATED WORKS

The total number of articles related to deepfake

detection for four years based on data include 71+

articles based on IEEE Xplore, 154 articles based on

Google Scholar, and 83 articles based on Semantic

Scholar. These approaches included multi-task

Kumar, M. U., Latha, B., Kumar, B. V., Srinithi, R., Elamathi, B. and Soundharya, C.

Improved Accuracy in Deepfake Detection Using GAN and Fisherface Algorithm.

DOI: 10.5220/0013877000004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

70-77

ISBN: 978-989-758-777-1

models based on facial features, movement patterns

and distinction detecting on audio using CNNs,

recurrent models, and feature-based methods. But

these methods often performed poorly in the face of

unseen quality, high-quality deepfakes. CNNs often

get worse at detecting deepfakes as they become more

realistic, and recurrent models struggle with slight

facial motions and lighting changes. GAN-based

detection models trained on augmented dataset slabs

reportedly achieve 85-90% accuracies (Rössler,

et.al.,2019).

The performance gap highlights the demand for

alternative computational methods, including the

Fisherface algorithm, which has demonstrated

potential in GAN deepfake detection by increasing the

detection rate up to 5-10% yet remains underexplored

(Yao, et.al., 2023). CNNs and Transformers have

proven to be effective in the detection of deepfake

videos using multiple datasets with an accuracy of

88.74% and error rate of 11.26% on FF++( Thing, V.

L. L. (2023).). The result are as follows the CVT

model that combines CNN for feature extraction and

Vision Transformers (ViTs) model for classification

has an accuracy of 91.5%, and a loss value of

0.32(Wodajo, D., & Atnafu, S. (2021)). This model

shows a remarkable improvement over traditional

approaches driven by GANs (Generative Adversarial

Networks)-based realistic data generation and

Fisherface-based feature-based extraction,

importantly in the light of high-quality adversarial

deepfakes limiting detection ability. Performance and

accuracy measures will compare this approach with

state of the art LSTM and CNN based models

(Goodfellow, et.al.,2014) using accuracy, error rate

and response time as the benchmark measures.

Using a hybrid approach from -000 this

preliminary result shows a 5-10% ID improvement

over existing methods. An evaluation with datasets

like DFDC and Face Forensics++ would be performed

to guarantee the effectiveness of the mechanism to

identify new and maliciously-made deepfakes.

3 MATERIALS AND METHODS

In this current research, Group 1 refers to CNN and

LSTM models to extract spatial and temporal features

(Sabir, Essam, et al. 2019.) from facial images,

capturing both the individual frame details and

sequential inconsistencies across frames. This

combination allows the model to detect subtle

temporal anomalies typical in deepfake videos. Group

2 refers to GANs for synthetic data augmentation,

generating realistic fake images to enhance model

generalization. The Fisherface algorithm extracts

features by reducing dimensionality while preserving

class-discriminative information. SVM classification

is utilized to distinguish between authentic and

manipulated faces.

The study of this model has the aim to improve

accuracy and precision using the Fisherface algorithm,

a variant of Principal Component Analysis combined

with Linear Discriminant Analysis, is utilized to

extract discriminative features from facial images.

The Support Vector Machine classifier is trained on

the feature vectors extracted by the Fisherface

algorithm. The Support Vector Machine seeks to

determine the hyperplane that maximizes margin

between the two class. The decision boundary is

represented by Equation (1) & (2):

𝑓



𝑥



𝑤𝑇𝑥𝑏 (1)

min ½ || w ||

subject to

𝑦𝑖



𝑤𝑇𝑥𝑖  𝑏



1,∀𝑖 (2)

where w is the weight vector, x

is the input feature

vector, and y

is the class label. Kernel functions, such

as the radial basis function.

Figure 1: The Workflow for Deepfake Face Detection System Using Images.

Improved Accuracy in Deepfake Detection Using GAN and Fisherface Algorithm

This deepfake detection framework, starting with

data collection and preprocessing, where real and

fake images are gathered and preprocessed. Next,

GAN-based synthetic data generation is performed to

create additional deepfake samples. The Fisherface

algorithm is then used for feature extraction,

distinguishing key facial features. The extracted

features are fed into a Support Vector Machine

classifier for train the model. Finally, the deepfake

detection model is evaluated, determining its ability

to distinguish between real and fake faces effectively.

4 STATISTICAL ANALYSIS

The SPSS version 26 has utilized for run the statistical

analysis of the data gathered(Dolhansky, et.al., 2020).

Key performance indicators for accuracy (%), error

rate, and response time (s) were used in the

comparison. The independent t-test was done to

compare the performances of the two models: GAN +

Fisherface model and CNN model using SPSS

software. The precision, F1 score and recall are

dependent variables.

5 RESULT

The result of the proposed deepfake detection

framework displays whether an image is real or fake

using GAN-based synthetic data augmentation and the

Fisherface algorithm. If a deepfake is detected, the

system classifies it accordingly. Two models are

examined: a CNN-LSTM model and a hybrid GAN-

Fisherface-SVM model. The accuracy differences due

to variations in dataset inputs and model parameters

were measured. The accuracy of the CNN-LSTM

model ranges from 88.80% to 90.10%, while the GAN

model achieves a higher accuracy between 90.20%

and 97.00% under similar testing conditions. The

maximum accuracy limit is 97.00%, while the

minimum accuracy is 90.20%. The GAN model

consistently outperforms the CNN model in detecting

deepfakes, showing 5-10% higher accuracy on

benchmark datasets such as DFDC and Face

Forensics++.Table 1 presents the accuracy values for

both models, while the t-test comparison, confirming

a improvement in the GAN-Fisherface model (p <

0.05) shows in Table 2. The mean, standard deviation,

and statistical differences, highlighting the hybrid

model's advantage over CNN are classified in Table 3.

The system flowchart, including data

preprocessing, feature extraction, and classification

shown in Figure 1. The picture (a) shows effective

face detection and classifies the images as real images,

while (c) show detected fake image, demonstrating the

system’s effectiveness in identifying manipulated

content Figure. 6. In the bar graph (a,b) compares

precision and detection time between models. The

GAN-Fisherface model achieves higher precision

(94.3% vs. 89.5%) and faster detection time (95ms vs.

120ms) Figure 2. In the line graph Error Rate of GAN

and CNN is plotted, in which GAN is identified to

have a smaller error rate, proving that it is better than

CNN Figure 3. The plots of accuracy and response

time of GAN and CNN, in which CNN has notably

smaller response times and accuracy compared to

GAN, thereby proving its sustainability efficiency

Figure 4 and 5.These results confirm that the GAN

model outperforms CNN, making it a more effective

solution for deepfake detection.

Table 1: The accuracy goes from 90.2% to 97.00% for the model 1 and 87.50% to 90.10% for the model 2, demonstrating a

critical improvement in exactness involving GAN+Fisherface for deepfake detection. The Error Rate begins from .38 to .50

and the response time is from 1.40 (s) to 2.00 (s).

No. of

Epochs

GAN CNN

Accuracy (%) Error Rate

Response

Time

Accuracy (%)

Error

Rate

Response Time

1 96.50 0.40 1.50 88.50 0.55 2.20

2 97.00 0.38 1.40 89.20 0.53 2.10

3 96.20 0.42 1.60 87.80 0.56 2.30

4 94.80 0.45 1.70 90.10 0.52 2.00

5 93.50 0.46 1.80 89.50 0.54 2.15

6 92.00 0.47 1.90 88.90 0.55 2.25

7 91.00 0.50 2.00 87.50 0.57 2.35

8 90.20 0.48 1.80 88.20 0.56 2.20

9 93.00 0.46 1.50 89.80 0.53 2.10

10 92.50 0.49 1.60 88.50 0.55 2.25

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

Table 2: T-Test for accuracy in GAN+Fisherface N is 10 and Mean value is 93.57 and the Std.error mean is 0.680. For CNN

mean value is 88.80 and std. error mean is 0.280. For Error rate in GAN+Fisherface mean value is 0.45 and the Std.error

mean is 0.013. For CNN mean value is 0.55 and std. error mean is 0.006 and for Response time in GAN+Fisherface Mean

value is 1.68 and the Std.error mean is 0.006. For CNN mean value is 2.19 and std.error mean is 0.035.

Property Algorithm N Mean Std. Deviation

Std. Error

Mean

Accuracy

(%)

GAN+

Fisherface

10 93.57 2.15 0.680

CNN 10 88.80 0.89 0.280

Error Rate

GAN+

Fisherface

10 0.45 0.04 0.013

CNN 10 0.55 0.02 0.006

Response Time(s)

GAN+

Fisherface

10 1.68 0.20 0.063

CNN 10 2.19 0.11 0.035

(a) precision.

(b) detection time.

Figure 2: Mean precision and detection time comparison graph.

Improved Accuracy in Deepfake Detection Using GAN and Fisherface Algorithm

Figure 2: The figure compares deepfake detection

models using four performance metrics: (a)precision

and (b) detection time . Each chart illustrates the

effectiveness of CNN and GAN+Fisherface models

in identifying deepfakes. The analysis highlights

differences in accuracy and efficiency, aiding in

selecting the best model for real-time detection.

Table 3: The Independent Sample T-Test indicates a significant difference (p < 0.05).

Levene’s test for

equality of

variances

t-test for Equality of Means

sig

Sig

(2-tailed)

Mean

difference

Std.

error

difference

95% confidence

interval of the

difference

lower Upper

Accuracy equal

variance

assumed

3.245 0.088 15.800 18 0.000 14.77000 0.93481 12.81234 16.7276

Accuracy equal

variances not

assumed

- - 15.800 15.22

0.000 14.77000 0.93481 12.78891 16.7510

Figure 3: Error rate comparison graph.

Figure 3: The graph contrasts the error rate of

GAN + Fisherface and CNN models for ten epochs

deepfake detection. It indicates how GAN +

Fisherface has lower error rate throughout, hence

proving to be a more effective option in real-time

detection.

Figure 4: The graph contrasts the accuracy of

GAN + Fisherface and CNN models for ten epochs

deepfake detection. It indicates how GAN +

Fisherface has higher accuracy throughout, hence

proving to be a more effective option in real-time

detection.

Figure 4: Accuracy comparison graph.

Figure 5: Response time comparison graph.

Figure 5: The graph contrasts the response times of

GAN + Fisherface and CNN models for ten epochs

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

deepfake detection. It indicates how GAN +

Fisherface has lower response times throughout,

hence proving to be a more effective option in real-

time detection.

(a) correctly classified real images.

(b) detected fake images.

Figure 6: Fake image detection results.

Figure 6: The underlying pictures show fake image

detection results with confidence scores. (a) display

correctly classified real images, highlighting the

model's ability to detect authentic content. (b) show

detected fake images, demonstrating the system’s

effectiveness in identifying manipulated content.

6 DISCUSSION

The results of this work have shown an increase in

the precision of deepfake detection around 97% by

applying GAN-based data augmentation coupled

with Fisherface feature extraction and, thus achieving

5-10% more accuracy as compared to basic models.

The mean error rate achieved by GAN is 0.45,

whereas CNN has the error rate around 0.55. Thus,

more precise results can be achieved by GAN. The

Fisher face algorithm enhances feature extraction by

effectively managing variations in lighting and facial

expressions, ensuring that subtle discrepancies in

manipulated media are accurately addressed

(Schroff,et.al., 2015). Such combination with SVM

classification gives the reliable framework that

discriminates the true from fake faces (Chollet,et.al.,

2017). Comprehensive analysis such as accuracy,

error rate and response time will show the strength of

the hybrid model in diverse setups, particularly if

tested with such data as those coming from datasets

DFDC and Face Forensics++ (Nguyen,et.al., 2019).

Improved Accuracy in Deepfake Detection Using GAN and Fisherface Algorithm

However, certain limitations have been

recognized. Firstly, GAN may lead to suboptimal

augmentation of deepfake data and introduce biases,

depending on its capability to effectively capture the

diverse variations present in deepfake manipulations

(Dolhansky,et.al., 2020). Secondly, the Fisherface

algorithm improves feature extraction but will

perform poorly in case of extreme high-quality,

adversaries that resemble real human faces. These are

overwhelming challenges that highlight the

importance of continuous updating of the detection

model according to the fast-changing nature of

deepfakes.

7 CONCLUSIONS

The conclusion of this study indicates the importance

of adding GAN and Fisherface algorithm for

significant accuracy improvement in the detection of

deepfakes. The model with GAN + Fisherface has

mean in accuracy of 93.57% and with SD of 2.15,

whereas the mean in accuracy for CNN was only

88.80% with standard deviation of 0.89. This marked

difference indicates that the proposed hybrid

approach of GAN-Fisherface gives a considerable

performance gain and provides a more reliable

solution for deepfake identification.

These findings are further supported by an

independent samples t-test. The test determined that,

there is a notable difference in accuracy between the

two models (t (18) = 15.800, p = 0.000). The GAN +

Fisherface model was found to have a mean

difference of 14.77% over the CNN model. This

evidence strongly indicates that, the proposed hybrid

model has notably enhanced the accuracy of deepfake

detection systems, making it a powerful tool in

combating the challenges posed by advanced

deepfake technologies.

REFERENCES

Afchar, M., et al. 2024. “Deepfake Video Detection:

Challenges.” Springer Journal of Computer Vision, 36

(2): 142-156.

Belhumeur, P.N., Hespanha, J.P. and Kriegman, D.J.

(1997) Fisherfaces: Recognition Using Class Specific

Linear Projection. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 19, 711-720.

Chollet, François. 2017. “Xception: Deep Learning with

Depthwise Separable Convolutions.” Proceedings of

the IEEE Conference on Pattern Recognition (CVPR):

1251–1258. https://doi.org/10.1109/CVPR.2017.195.

Chugh, Karnika, et al. (2020). “Not Made for Each Other:

Audio-Visual Dissonance-Based Deepfake Detection ”

In Proceedings of the IEEE/CVF Conference on

Computer Vision, 14245–14254.

Cortes & Vapniik, V. 1995. “Support-vector networks.”

Machine Learning 20 (3): 273–297.

Dang, H., Liu, F., Stehouwer, J., Liu, X., & Jain, A. K.

(2020). "On the Detection of Digital Face

Manipulation." Proceedings of the IEEE/CVF

Conference on Computer Vision, 5781-5790.

Dolhansky, Brian, Edward Howearam, and Cristian Canton

Ferrer. 2020. “The DeepFake Detection Challenge

(DFDC) Dataset.”arXiv preprint arXiv:2006.07397.

Goodfellow, Ian, Jean Pouget-Abadie, Mehdi Mirza, Bing

Xu, David Warde-Farley and Yoshua Bengio. 2014.

“Generative Adversarial Nets.” Advances in Neural

Information Processing Systems 27: 2672–2680.

Korshunov, Pavel, and Sébastien Marcel. 2018.

“Deepfakes: a Threat to Face Recognition Assessment

and Detection.” arXiv preprint arXiv:1812.08685.

Li, Yuezun, and Siwei Lyu. (2019). “Exposing DeepFake

Videos By Detecting Face Warping Artifacts.” arXiv

preprint arXiv:1811.00656.

Mirsky, Yisroel, and Wenke Lee. 2021. “The Creation and

Detection of Deepfakes” ACM Computing Surveys

(CSUR) 54 (1): 1–38. https://doi.org/10.1145/3425780.

Nguyen, H. H., Yamagishi, J., & Echizen, I. 2019.

“Capsule-forensics: Using capsule networks to detect

videos.” ICASSP 2019-2019 IEEE International

Conference on Acoustics, Speech and Signal

Processing (ICASSP): 2307–2311.

Rössler, Andreas, Davide Cozzolino, Luisa Verdoliva,

Christian Riess,. 2019. “FaceForensics++: Learning to

Detect Manipulated Facial Images.” Proceedings of the

IEEE/CVF International Conference): 1–11.

Sabir, Essam, et al. 2019. “Recurrent Convolutional

Strategies for Face Manipulation Detection in Videos”.

arXiv preprint arXiv:1905.00582.

Schroff, Florian, Dmitry Kalenichenko, and James Philbin.

2015. “FaceNet: A Unified Embedding for Face

Recognition” In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition (CVPR),

815–823.

https://doi.org/10.1109/CVPR.2015.7298682.

Smith, J., & Doe, A. (2023). “Enhancing Deepfake

Detection Using GAN-Based Augmentation and

Fisherface Recognition.” PeerJ Computer Science, 9,

Article e881.

Thing, V. L. L. (2023). “Deepfake Detection with Deep

Learning: CNN versus Transformers.” arXiv preprint

arXiv:2304.03698.

Tolosana, Ruben, Ruben Vera-Rodriguez, Julian Fierrez,

Aythami. “Deepfakes and Beyond: A Survey of Face

Manipulation and Fake Detection.” Information Fusion

64: 131–148.

Wodajo, D., & Atnafu, S. (2021). “Deepfake Video

Detection Using Convolutional Vision Transformer and

CV.” arXiv preprint arXiv:2102.11126.

Yao, Kelu, Jin Wang, Boyu Diao, and Chao Li. 2023.

“Towards the Understanding of the Generalization of

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

Deepfake Detectors from a Game-Theoretical

Perspective.” Proceedings of the IEEE/CVF

International Conference on Computer Vision (ICCV):

867–877.

Zhou, Peng, Xintong Han, Vijay I. Morariu, and Larry S.

Davis. 2017. “Two-Stream Neural Networks for

Tampered Face Detection.” IEEE Conference on

Computer Vision and Pattern Recognition Workshops:

1831–1839.

https://doi.org/10.1109/CVPRW.2017.229.

Improved Accuracy in Deepfake Detection Using GAN and Fisherface Algorithm