Enhanced Deepfake Detection Using ResNet50 and Facial Landmark

Analysis

P. S. Prakash Kumar

, G. Mohesh Kumar

, K. Shanmuga Priya

, V. Dharun Kumar

P. Nanda Krishnan

and S. A. Barathraj

Department of Information Technology, K S R College of Engineering, Tiruchengode 637215, Tamil Nadu, India

Department of Information Technology, K S R Institute for Engineering and Technology, Tiruchengode‑637215, Tamil

Nadu, India

Keywords: Deepfake Detection, ResNet50, Deep Learning, MobileNetV2, Detection Accuracy, Fake Image, Dataset,

Face.

Abstract: This research focuses on accuracy enhancement in the detection of deepfakes using the ResNet50 algorithm

designed through deep learning. It analyzes anomalies in artificial facial images. Materials and Methods: The

two implemented deep learning models include MobileNetV2 (Group 1) and ResNet50 (Group 2), each

trained and tested with 40 image samples, comprising 20 real images and 20 deepfake images. Here, a facial

irregularity detector based on ResNet50 was trained against one whose model was created through

MobileNetV2. Result: ResNet50 was shown to have a detection accuracy of 91.81 % to 97.87 % for

distinguishing between real and fake photographs. Its effectiveness for real-time applications is demonstrated.

Statistical study revealed a significant improvement in detection accuracy than the MobileNetV2 Model (p-

value < 0.05). Conclusion: According to the study's results, the ResNet50 algorithm is very good at identifying

deepfake photos and real photos with a low mistake rate and high accuracy. Due to its efficiency in processing

synthetic and genuine images, it can be a dependable tool for handling the problems created by deepfake

media.

1 INTRODUCTION

This research focuses on identifying deepfakes using

Residual Neural Networks (ResNet50), leveraging

their proven success in image recognition (D. Zhu,

2025) The study examines the detection performance

of ResNet50 in distinguishing deepfake images by

extracting subtle irregularities and inconsistencies

introduced during their generation (C. Yang, 2025)

Deep learning advancements have improved methods

of detecting deepfakes, and the project here is to

improve this method using the robust and reliable

architecture of ResNet50. Training ResNet50 on a

large dataset of real and manipulated images/videos

would enable the recognition of anomalies that

distinguish deepfakes from authentic content (S.

Xue,2025) The objective of this research is to use a

deep learning-based ResNet50 architecture to develop

a trustworthy deepfake detection system. The

ResNet50 model will be trained on a large dataset of

actual and changed images or videos, allowing the

system to identify subtle irregularities and

inconsistencies introduced throughout the deepfake

manufacturing process. This research is important

since deepfakes seriously jeopardize the legitimacy of

digital information and affect journalism, media, and

law enforcement (S. Song, 2025) The strategy will

increase public trust in digital media and assist stop

the spread of false information (A. Qadir, et.al,2024),

(R. Yang, et.al, 2024), (S. M. Hassan, et.al,2025), (M.

Saratha et al,2025).

2 RELATED WORKS

The total number of articles published on this topic in

the last five years, there have been more than 1954

publications on the topic in IEEE Xplore and 836 in

Google Scholar. In contrast to conventional

techniques, the combination of preprocessing steps,

feature-based, residual connection, and batch

normalization improves the detection accuracy of

deepfake videos in the presence of facemasks.

848

Prakash Kumar, P. S., MoheshKumar, G., Shanmuga Priya, K., Dharun Kumar, V., Nanda Krishnan, P. and Barathraj, S. A.

Enhanced Deepfake Detection Using ResNet50 and Facial Landmark Analysis.

DOI: 10.5220/0013948000004919

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 5, pages

848-855

ISBN: 978-989-758-777-1

According to the study's findings, face-mask-

deepfakes can be detected with 94.81 % accuracy

when compared to the conventional

InceptionResNet50V2 and VGG19, which have

accuracy rates of 77.48 % respectively. Future

research should assess the precision of creating a

follow-up experimental study for improved deepfake

detection using facemasks.

The user interface further displays a confidence

score for each prediction to provide users an idea of

the model's dependability. With training and testing

accuracy rates over 95 %, this study provides a

comprehensive technique to differentiate between real

and fake faces (C. Wang, 2025) The ability of deep

learning to produce and identify deepfakes is still

developing. deepfake detection models are created

using older datasets, may become obsolete over time,

and constantly need new detection methods. With an

accuracy rate of over 90 %, the research findings are

encouraging, but there is still room for improvement

(N. M. Alnaim, et.al 2025) (S. Karthikeyan et

al,2020)The CT outperforms the state-of-the-art in the

deepfake detection task, demonstrating its better

discriminative capacity and resilience to various forms

of attacks. With an overall classification accuracy of

over 97 % despite accounting for deepfakes from 10

different GAN architectures that are not only involved

in face photos, the CT demonstrates its dependability

and independence from image semantics. Finally, tests

employing deepfakes created by FACEAPP

demonstrated the effectiveness of the proposed

strategy in a real-world scenario, achieving 93 %

accuracy in the fake detection task (Y. Xu, 2025)

Additionally, in order to assess the resilience of

Arabic-AD, Arabic recordings from non-Arabic

speakers were gathered, taking into account the accent.

To evaluate the suggested approach and contrast it

with established standards in the literature, three

comprehensive experiments were carried out. With

the lowest EER rate (0.027 %) and the highest

detection accuracy (97 %) while avoiding the

requirement for lengthy training, Arabic-AD thereby

surpassed other state-of-the-art techniques (L. Zhou,

2025) The study's YOLO models all exhibited almost

flawless accuracy in recognizing X-ray images that

revealed osteoarthritis in the knee. Despite the

generally excellent performance of the YoloV8

models, the YoloV5 models yielded the best and

poorest results in lung CT scan images.The YoloV5su

model had the highest recall (0.997), while the

YoloV5nu model had the lowest (0.91). Additionally,

YoloV5su, the best model, performs 60 % better than

YoloV8x, the second-best model. The results

demonstrate YoloV5su's speed and accuracy in

detecting medical deepfakes (T. Qiao, 2025) (P.

Jahnavi et al, 2025) The neural network will be trained

extensively using these photos and their labels in order

to accurately predict an image's legitimacy. The model

has been trained on precisely 7104 photos, and based

on the outcomes, the model's predictions are

reasonably accurate. The accuracy and precision of the

results are 88.81 % and 87.93 %, respectively. The

suggested technique classifies the input video frames

into genuine and deepfake by extracting features using

a CNN-MLP model that is based on deep learning.

The suggested approach detected deepfake movies

with a high accuracy of 81.25 % using Celeb-df, a

publicly available deepfake dataset. Comparative tests

using a different algorithm show how accurate our

model is, confirming its effectiveness in identifying

modified content (N. Saravanan et al, 2025) The

feature extraction stage, which is the next and most

important step, is when particular features relating to

face orientation are taken out of every frame. A

histogram of oriented gradients, scale-invariant

feature transforms, and local binary patterns are the

features that are extracted (N. K. Wardana, 2022).

Normalization of features takes place in other neural

network pipelines later in the design. The following

neural network will then categorize the videos as

either deepfake or not after combining the features.

This architecture has a validation accuracy of 95.89 %

(M. Venkatesan et al,2025)

Previous studies have shown that using

MobileNetV2 methods to detect deepfake content

might be challenging. An important factor in the

rapidly changing realm of cybersecurity is effective

deepfake detection. In contrast to traditional detection

methods, the goal of this study is to improve detection

accuracy by employing a unique deep learning-based

ResNet50 algorithm.

3 MATERIALS AND METHODS

The study measures the performance of the ResNet50

model compared with that of MobileNetV2 in

deepfake image detection based on predictive

accuracy. The models are built and tested on a

prepared dataset to verify real vs fake image

classification (L. Pham, 2025) The experiments were

conducted in the IT Lab at KSR Institute for

Engineering and Technology, where high-

performance computing systems and necessary

computational support are available for deep learning-

based research. The dataset used for training and

testing was obtained from Kaggle.com, specifically

the Deepfake Detection Dataset. Statistical methods

Enhanced Deepfake Detection Using ResNet50 and Facial Landmark Analysis

849

such as a 95 % confidence interval and a 0.05 % level

of significance at a G-power value of 80 % were used

to validate the results.

In this study, Group 1 is the MobileNetV2 model,

which was trained and tested on 2178 deepfake and

real images. The model, however, failed to predict

images well, resulting in inconsistent performance in

detection and obtained an accuracy level of 89 % (E.

Şafak, 2017) Group 2 refers to the ResNet50 deep

learning model, trained and tested on 2193 deepfake

and real images. The aim is to improve detection

accuracy through deeper feature extraction and

enhanced classification capabilities. The ResNet50

model was found to have improved the detection of

deepfake images by about 5 % to 7 % when compared

to MobileNetV2.

The Deepfake detection system begins with the

collection of the dataset from sources such as Kaggle,

and preprocessing steps like resizing, normalization,

and augmentation are performed on it. Then, the data

is split into training and testing.

Figure 1: The workflow of the deepfake detection process

using images.

The ResNet50 model learns deepfake-specific

patterns during the training phase. Feature extraction

and facial landmark analysis using dlib enhance

classification accuracy. The figure 1 shows the

workflow of the deepfake detection process using

images Metrics like accuracy, precision, recall, and

F1-score are used to evaluate the trained model.

During detection, the final phase, this model classifies

images as being real or fake based on some

confidence scores which ensure an efficient

deepfake-detecting model.

4 STATISTICAL ANALYSIS

The presentation of the deepfake recognition

framework is analyzed by using SPSS adaptation 26

for a quantitative study. The review centers on two

essential measurements: discovery precision and

levels of certainty (K. Stehlik-Barry, 2017) A one-tail

free t-test was conducted to analyze the results. The

ResNet model showed high dependability,

accomplishing more than 97 % exactness for ensured

pictures and 94 % accuracy for counterfeit ones. These

outcomes highlight the design's capacity to see

genuine and controlled pictures with surprising

accuracy, making it a strong reaction for combating

deepfakes.

5 RESULT

The results show that ResNet50 is more accurate,

with accuracy ranging from 91.81 % to 97.87 %,

while for the MobileNetV2 variants it was between

83.64 % and 89.56 %. It follows that the error for

ResNet50 lies in the range of 2.13 % to 8.19 %

compared to a much higher range for MobileNetV2,

which is from 10.44 % to 13.71 %. Further statistical

measures such as standard deviation, variance, and t-

tests emphasize the significance of the differences.

Graphical analysis further emphasizes the better

nature of ResNet50 about its always good

performance.

The T-test comparison between the ResNet50

and MobilseNetV2 models shows a high difference in

the accuracy, with p < 0.05. ResNet50 has higher

mean accuracy as 95.155 with low standard deviation

0.86375, and MobileNetV2 has mean accuracy as

87.557 with a high standard deviation of 1.01267.

This indicates that although ResNet50 is more

accurate, its consistency is slightly higher due to

lower standard deviation. The graph of accuracy

distribution also shows the comparative advantage of

ResNet50: multiple iterations indicate higher

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

850

accuracy in the case of the ResNet50 model. This

again shows that the model proposed here is better

and more consistent than MobileNetV2 in terms of

deepfake detection.

Table 1: The model of MobileNetV2 realizes an accuracy of 86.37 % to 89.56 %, an error rate of 10.87 to 13.63.

Iteration Accuracy Error Rate

1 86.37 13.63

2 87.29 12.71

3 86.72 13.28

4 88.02 11.98

5 88.64 11.36

6 89.56 10.44

7 86.42 13.58

8 89.13 10.87

9 87.81 12.19

10 86.38 13.62

11 86.44 13.56

12 87.31 12.69

13 88.14 11.86

14 86.69 13.31

15 87.55 12.45

16 86.71 13.29

17 89.05 10.95

18 87.86 12.14

19 86.74 13.26

20 88.32 12.14

Table 2: The model of ResNet50 realizes an accuracy of 91.81% to 97.87% an error rate of 2.13 to 7.65.

Iteration Accuracy Error Rate

1 94.24 5.76

2 95.41 4.59

3 96.05 3.95

4 97.12 2.88

5 96.80 3.20

6 97.35 2.65

7 95.85 4.15

8 97.87 2.13

9 96.30 3.70

10 94.75 5.25

11 96.51 3.49

12 91.81 8.19

13 92.48 7.52

14 94.03 5.97

15 92.35 7.65

16 94.92 5.08

17 93.49 6.51

18 94.95 5.05

19 94.63 5.37

20 95.46 6.54

Enhanced Deepfake Detection Using ResNet50 and Facial Landmark Analysis

851

Table 3: The mean Accuracy of ResNet 50 is 95.155 and MobileNetV2 has 87.557. Even though ResNet50 has a significantly

lower variability or std. Deviation of 0.86375 while MobileNet v2 has Std. Deviation of 1.01267, means only slightly high

stability.

Metric Method N Mean Std. Deviation Std. Error Mean

Accuracy MobileNetV2 20 87.557 1.01267 0.22644

Accuracy ResNet50 20 95.155 0.86375 0.15378

Table 4: Form SPSS Independent Samples Test t-test Comparison of gain in MobileNet v2 and ResNet 50 (p<0.05).

Levene’s test

for equality of

variances

Independent samples test

sig

Sig

(2-

tailed

)

Mean

difference

Std.

error

difference

95% confidence

interval of the

difference

lowe

Accuracy

equal

variance

assume

4.381 0.043

16.761

38 0.000 -7.59800 0.45332

8.51570

6.68030

Accuracy

equal

variances

not

assume

16.761

30.377 0.000 -7.59800 0.45332

8.52332

6.67268

Figure 2: Accuracy comparison between MobileNetV2 and

ResNet50.

Figure 3: Accuracy in the classification.

Figure 2, the outcome with the Accuracy is

displayed along with the result of the uploaded image,

and it determines that the image is real with high

Accuracy as illustrated in 2 and identifies the image

as fake with strong Accuracy in the classification as

shown in figure 3. The interface clearly displays the

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

852

outcome of the classification to ensure an efficient

process of deep fake detection.

Figure 4: Accuracy comparison.

The accuracy comparison between MobileNetV2

and ResNet50 is shown in Figure 4 and while the error

rate is depicted in Figure5. The performance of

ResNet50 was better compared to both the metrics,

that is, with greater accuracy and maintaining a lesser

error rate compared to MobileNetV2 throughout all

iterations.

Figure 5: Error rate comparison.

Figure 6: Accuracy comparison.

Figure 7: Error rate comparison.

Figure 6 contrasts the accuracy of MobileNetV2

and ResNet50 with a bar graph, and Figure 7 uses

another bar graph to depict their error rates. In all

iterations, ResNet50 outperforms MobileNetV2 by

having greater accuracy and lower error rates, which

makes them highly effective in detecting DeepFakes.

6 DISCUSSION

The present study showed how the ResNet50

algorithm provides a remarkable degree of

improvement in detecting deepfake images compared

to the MobileNetV2 model (L. Si et al, 2025) With

accuracies between 91.81 and 97.87 % achieved with

ResNet50, the model took precedence over

MobileNetV2, showing a significance in difference

with statistical confidence (p-value < 0.05) (S.

Khairnar,2025) (P. Raghavendra Reddy et al, 2025)

This emphasizes the potential of ResNet50 for

accurately distinguishing between real and deepfake

images, especially in real-time applications where

accuracy becomes critical. With skip connections

incorporated into a deep architecture, ResNet50 is

capable of extracting rather complex features from

facial images that sensitively detect small anomalies

which might become possible red flags in deepfakes

but invisible to standard detection methods (Y.

Mao,205) The model can also learn hierarchical

representations of features across many layers that

further augment its power to identify variations in

textures and contours and minor details that remain

undetectable for standard economic functions. For

example, slight differences in skin texture, eye

movements, or maybe facial expressions are

manipulations that happen in deep-fake content (M.

C. Gursesli, 2025) (A. V. Santhosh Babu et al, 2019)

Slightest differences are what make ResNet50 best

Enhanced Deepfake Detection Using ResNet50 and Facial Landmark Analysis

853

capable of doing deep-fake detection and, therefore,

good at unveiling minute manipulations

(M.Venkateswarlu, 2025) (N.Saravanan et al,2025)

Deepfake generation methods, including facial

swapping, expression manipulation, or synthesis-

based facial animation, are all variable methods that

will require a contrasting range of capabilities for

accurate identification and distinguishable traits

incorporated within them. ResNet50 scores high in

feature extraction, which allows it to withstand

variation in the described attributes (H. Chen, 2025)

Due to its great generalizability across varied

datasets, it can distinguish deepfakes coming from

different sources and formats, which is an exciting

feature for real-world deployment. The extremely

high accuracy shows that deep learning-based

methods have prospects like ResNet50. Future work

will extend to increase diversity of the dataset,

enhance pre-processing, and boost the accuracy with

ensemble learning. Attention mechanisms and

adversarial training will improve robustness.

Optimizations will minimize further computational

overhead. Extracting both image and audio features

for generating the model for detection shall increase

this performance.

7 CONCLUSIONS

The deepfake detection system was designed to train,

and test for the performance in the differentiating

between real and deepfake images utilizing the

ResNet50 algorithm. The results showed that the

ResNet50 model displayed significantly better than

the performance of the MobileNetV2 model and

demonstrated between 91.81 % to 97.87 % detection

accuracy, while MobileNetV2 achieved 86.37 % to

89.56 % accuracy. This signifies the wide margin for

the improvement of the determines ResNet50 model's

capability of detecting minute anomalies and

manipulations in facial images and essential for high-

accuracy applications and considering stability with

the MobileNetV2 showing a standard deviation of

1.01267, indicating some variability in the

performance across the different datasets. On the

other hand, the lower standard deviation of 0.86375

for the ResNet50 model indicates the model had

consistently good performance across varying

datasets.

REFERENCES

A. V. Santhosh Babu et al, “Performance analysis on

cluster-based intrusion detection techniques for energy

efficient and secured data communication in MANET,”

International Journal of Information Systems and

Change Management, Aug. 2019, Accessed: Feb. 03,

2025. [Online]. Available: https://www.inderscienceo

nline.com/doi/10.1504/IJISCM.2019.101649

A. Qadir, R. Mahum, M. A. El-Meligy, A. E. Ragab, A.

AlSalman, and M. Awais, “An efficient deepfake video

detection using robust deep learning,” Heliyon, vol. 10,

no. 5, p. e25757, Mar. 2024.

C. Yang, S. Ding, and G. Zhou, “Wind turbine blade

damage detection based on acoustic signals,” Sci Rep,

vol. 15, no. 1, p. 3930, Jan. 2025.

C. Wang, C. Shi, S. Wang, Z. Xia, and B. Ma, “Dual-Task

Mutual Learning with QPHFM Watermarking for

Deepfake Detection.” Accessed: Feb. 03, 2025.

[Online]. Available:

https://doi.org/10.1109/LSP.2024.3438101

D. Zhu, C. Li, Y. Ao, Y. Zhang, and J. Xu, “Position

detection of elements in off-axis three-mirror space

optical system based on ResNet50 and LSTM,” Opt

Express, vol. 33, no. 1, pp. 592–603, Jan. 2025.

E. Şafak and N. Barışçı, “Detection of fake face images

using lightweight convolutional neural networks with

stacking ensemble learning method,” PeerJ Comput

Sci, vol. 10, p. e2103, Jun. 2024.

H. Chen, G. Hu, Z. Lei, Y. Chen, N. M. Robertson, and S.

Z. Li, “Attention-Based Two-Stream Convolutional

Networks for Face Spoofing Detection.” Accessed:

Feb. 03, 2025. [Online]. Available: https://doi.org/10.

1109/TIFS.2019.2922241

I. N. K. Wardana, “Design of mobile robot navigation

controller using neuro-fuzzy logic system,” Computers

and Electrical Engineering, vol. 101, p. 108044, Jul.

2022.

K. Stehlik-Barry and A. J. Babinec, Data Analysis with

IBM SPSS Statistics. Packt Publishing Ltd, 2017.

L. Pham, P. Lam, T. Nguyen, H. Nguyen, and A. Schindler,

“Deepfake Audio Detection Using Spectrogram-based

Feature and Ensemble of Deep Learning Models.”

Accessed: Feb. 03, 2025. [Online]. Available:

https://doi.org/10.1109/IS262782.2024.10704095

L. Si et al., “A Novel Coal-Gangue Recognition Method for

Top Coal Caving Face Based on IALO-VMD and

Improved MobileNetV2 Network.” Accessed: Feb. 03,

2025. [Online]. Available: https://doi.org/10.1109/TI

M.2023.3316250

L. Zhou, C. Ma, Z. Wang, Y. Zhang, X. Shi, and L. Wu,

“Robust Frame-Level Detection for Deepfake Videos

with Lightweight Bayesian Inference Weighting.”

Accessed: Feb. 03,2025.

[Online].Available:https://doi.org/10.1109/JIOT.2023.

3337128

M. C. Gursesli, S. Lombardi, M. Duradoni, L. Bocchi, A.

Guazzini, and A. Lanata, “Facial Emotion Recognition

(FER) Through Custom Lightweight CNN Model:

Performance Evaluation in Public Datasets.” Accessed:

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

854

Feb. 03, 2025. [Online].Available:https://doi.org/10.1

109/ACCESS.2024.3380847

M. Venkatesan et al, “A New Data Hiding Scheme with

Quality Control for Binary Images Using Block Parity.”

Accessed: Feb.03,2025. [Online]. Available: https://d

oi.org/10.1109/IAS.2007.26

M. Saratha et al, “Research and Application of Boundary

Optimization Algorithm of Forest Resource Vector

Data Based on Convolutional Neural Network.”

Accessed: Feb. 03, 2025. [Online]. Available:

https://doi.org/10.1109/SmartTechCon57526.2023.103

91589

M.Venkateswarlu and V. R. R. Ch, “DrowsyDetectNet:

Driver Drowsiness Detection Using Lightweight CNN

With Limited Training Data.” Accessed: Feb. 03, 2025.

[Online]. Available: https://doi.org/10.1109/ACCESS.

2024.3440585

N. Saravanan et al, “Revolutionizing Air Quality

Prognostication: Fusion of Deep Learning and Density-

Based Spatial Clustering of Applications with Noise for

Enhanced Pollution Prediction.” Accessed: Feb. 03,

2025. [Online]. Available: https://doi.org/10.1109/ICP

CSN62568.2024.00029

N. M. Alnaim, Z. M. Almutairi, M. S. Alsuwat, H. H.

Alalawi, A. Alshobaili, and F. S. Alenezi, “DFFMD: A

Deepfake Face Mask Dataset for Infectious Disease Era

with Deepfake Detection Algorithms.” Accessed: Feb.

03, 2025. [Online]. Available: https://doi.org/10.1109/

ACCESS.2023.3246661

N. Saravanan et al, “Accurate Prediction and Detection of

Suicidal Risk using Random Forest Algorithm.”

Accessed: Feb. 03, 2025. [Online]. Available:

https://doi.org/10.1109/ICPCSN62568.2024.00053

P. Raghavendra Reddy et al, “Novel Detection of Forest

Fire using Temperature and Carbon Dioxide Sensors

with Improved Accuracy in Comparison between two

Different Zones.” Accessed: Feb. 03, 2025. [Online].

Available: https://doi.org/10.1109/ICIEM54221.2022.

9853107

P. Jahnavi et al, “IOT based Innovative Irrigation using

Adaptive Cuckoo Search Algorithm comparison with

the State of Art Drip Irrigation to attain Efficient

Irrigation.” Accessed: Feb.03,2025. [Online].

Available: https://doi.org/10.1109/ICTACS56270.202

2.9987981

R. Yang, K. You, C. Pang, X. Luo, and R. Lan, “CSTAN:

A Deepfake Detection Network with CST Attention for

Superior Generalization,” Sensors (Basel), vol. 24, no.

22, Nov. 2024, doi: 10.3390/s24227101.

S. Karthikeyan et al, “An attempt to enhance the time of

reply for web service composition with QoS,”

International Journal of Enterprise Network

Management, Dec. 2020, Accessed: Feb.03,2025.

[Online]. Available: https://www.inderscienceonline.c

om/doi/10.1504/IJENM.2020.111750

S. M. Hassan and A. K. Maji, “Pest Identification Based on

Fusion of Self-Attention with ResNet.” Accessed: Feb.

03, 2025. [Online].Available:https://doi.org/10.1109/

ACCESS.2024.3351003

S. Song, J. C. K. Lam, Y. Han, and V. O. K. Li, “ResNet-

LSTM for Real-Time PM2.5 and PM₁₀ Estimation

Using Sequential Smartphone Images.” Accessed: Feb.

03, 2025. [Online]. Available:

https://doi.org/10.1109/ACCESS.2020.3042278

S. Khairnar, S. Gite, K. Mahajan, B. Pradhan, A. Alamri,

and S. D. Thepade, “Advanced Techniques for

Biometric Authentication: Leveraging Deep Learning

and Explainable AI.” Accessed: Feb. 03, 2025.

[Online]. Available: https://doi.org/10.1109/ACCESS.

2024.3474690

S. Xue and C. Abhayaratne, “Region-of-Interest Aware 3D

ResNet for Classification of COVID-19 Chest

Computerised Tomography Scans.” Accessed: Feb. 03,

2025. [Online]. Available: https://doi.org/10.1109/AC

CESS.2023.3260632

T. Qiao, S. Xie, Y. Chen, F. Retraint, and X. Luo, “Fully

Unsupervised Deepfake Video Detection Via Enhanced

Contrastive Learning.” Accessed: Feb.03,2025.

[Online].Available:https://doi.org/10.1109/TPAMI.20

24.3356814

Y. Xu, P. Terhörst, M. Pedersen, and K. Raja, “Analyzing

Fairness in Deepfake Detection with Massively

Annotated Databases.” Accessed: Feb. 03, 2025.

[Online]. Available: https://doi.org/10.1109/TTS.2024

.3365421

Y. Mao, Y. Lv, G. Zhang, and X. Gui, “Exploring

Transformer for Face Mask Detection.” Accessed: Feb.

03, 2025. [Online]. Available: https://doi.org/10.1109/

ACCESS.2024.3449802

Enhanced Deepfake Detection Using ResNet50 and Facial Landmark Analysis

855