INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
Development Of Artificial Intelligence-Based Model for Forensic  
Analysis of Cross-Platform Deepfakes  
*Joseph, C. C1, Ojeniyi, J. A2, Noel, M. D3, Ahmad, S4, Fasola, O. O5, Uduimoh, A6  
Department of Cyber Security Science, Federal University of Technology, Minna, Nigeria  
DOI: https://doi.org/10.51244/IJRSI.2025.1210000285  
Received: 31 October 2025; Accepted: 07 November 2025; Published: 19 November 2025  
ABSTRACT  
Synthetic media, a product of advancements in artificial intelligence (AI) and machine learning, represents a  
transformative innovation that reshapes how content is created, manipulated and consumed. One of the most  
advanced technology in synthetic media is deepfakes. The misuse of deepfakes poses significant threats to  
privacy, security and societal trust. Cybercriminals exploit this technology for phishing scams, identity theft  
and spreading misinformation. This study developed an AI-based hybrid model for forensic analysis of cross  
platform deep fakes to address these gaps. The study developed a hybrid model that integrates 2D CNN, 1D  
CNN and RNN framework capable of isolating platform-specific spatial and temporal features for forensic  
analysis of cross-platform deepfakes. The platforms include social media such as Facebook, Instagram,  
Youtube, Tiktok and Twiter. The research used Celeb DF dataset for training, validation and testing of the  
hybrid model. Result from the evaluation metrics showed that the model achieved an accuracy of 99 % on the  
training data and 93.5 % on the test data. Result equally showed that precision, recall and MCC value were  
96.8 %, 90 % and 87.2 % respectively on the test data. The model outperformed single CNN and RNN models  
and some hybrid models reported. This study demonstrated a reliable hybrid model with high detection  
accuracy for the forensic analysis of deepfakes. Hence, the model has the potential to address the problem of  
misinformation caused by synthetic media.  
Keywords: Artificial Intelligence, forensic analysis, 2D CNN, RNN, Deepfakes, Hybrid model  
INTRODUCTION  
In recent years, there has been a significant increase in the creation and dissemination of deepfake videos  
across various platforms [1,2]. This creation and dissemination of deepfake videos constituted both detection  
and security threats considering that deepfakes are manipulated videos from sophisticated deep learning  
techniques including Autoencoders and Generative Adversarial Networks (GANs) to produce contents that  
appear authentic but fabricated [3,4]. These deepfakes pose significant threats to digital authenticity, individual  
privacy, and public trust. Although, video manipulation is not a new phenomenon, however, the use of  
advanced technologies such as deep learning has revolutionised the process [5]. The misuse of deep learning in  
misinformation campaigns, cybercrimes and blackmail has raised global concerns. For instance, deepfakes are  
used to impersonate high-profile individuals. Typical example is a 2018 viral deepfake video featuring former  
President Barack Obama insulting President Donald Trump and making other controversial statements [3].  
Additionally, cybercriminals have exploited deepfakes to impersonate and defraud individuals and corporate  
organisations, a typical case is a 2024 video conference where an employee was tricked into transferring $25  
million during team meeting [6].  
Digital forensics involves the collection, verification and analysis of digital evidence in a manner that adhere to  
both legal and scientific standards [7]. Forensic tools and techniques must undergo peer review and be  
scientifically validated [8]. Large Language Models (LLMs), however, present challenges due to their lack of  
transparency regarding training data and internal processes, complicating the verification of their outputs'  
scientific reliability. Additionally, data confidentiality is critical in digital forensics and reliance on cloud-based  
Page 3275  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
LLMs raise concerns about exposing sensitive case information to third parties [9]. The role of digital evidence  
in law enforcement investigations has gained significant attention. This heightened focus stems from the  
pervasive nature of digital data in incidents requiring police scrutiny, as well as its capacity to provide detailed  
insights into the actions of involved parties. In the context of deepfake forensics, this trend has shown growing  
need for sophisticated digital forensic techniques to authenticate and analyse potentially manipulated media.  
As deepfakes become more prevalent and sophisticated, law enforcement agencies increasingly rely on  
advanced digital forensic methods to discern genuine content from artificial creations [10]. This is to ensure  
the integrity of evidence in criminal investigations  
Despite advancements in detection methods, existing forensic tools require complex models to address the  
challenges posed by differences in how videos or images are saved and compressed across devices and  
applications [11]. Most detection methods rely heavily on few metrics, but this study explored quite a number  
of metrics including the recent underutilised metric like the Matthews Correlation Coefficient (MCC) which  
are essential for evaluating forensic reliability in real world applications [12]. Furthermore, existing models are  
static and lack adaptability to evolving deepfake techniques and platform updates, which lead to rapid  
obsolescence  
[13].  
This study developed an AI-based hybrid model for forensic analysis of cross platform deep fakes to address  
these gaps. The model aims at enhancing detection accuracy by utilising hybrid 2D CNN, 1D CNN and RNN  
framework capable of isolating platform-specific artifacts. It also incorporated comprehensive evaluation  
metrics such as Accuracy, Area under the Curve (AUC), Precision, Recall, F1 Score, Specificity and the  
Matthews Correlation Coefficient (MCC). These metrics were chosen to ensure a balanced evaluation covering  
accuracy, correctness and practical usability. This addressed the existing limitations and provided a scalable  
solution for combating the growing threat of deepfakes across digital ecosystems.  
Related Work  
2.1 Customized deepfakes detection models  
Customized deepfakes detection models are AI systems tailored to identify manipulated images or videos with  
higher accuracy than generic models. They achieve this through: architecture customization, data-driven  
tailoring, and training techniques. One of such studies is the development of M-Task-SS (customised CNN)  
deepfake detection model that uses multi-task learning, self-supervision and incorporated additional layers  
such as dense layers, Max Pooling layers, and dropout layers to enhance detection accuracy to improve  
generalisation across different datasets [14]. Kosarkar et al. [15] explores the detection of deepfake images  
and videos using a customised Convolutional Neural Network (CNN) algorithm and compares its performance  
against two other CNN models. The customised CNN model, which includes additional layers such as a dense  
layer, Max Pooling, and a dropout layer. Almestekawy et al. [16] introduces a novel deepfake detection method  
focused on enhancing model stability and ensuring reproducible results, which are often overlooked in existing  
research. The technique combines multiple spatiotemporal textures with deep learning-based features, utilising  
an enhanced 3D Convolutional Neural Network (CNN) within a Siamese architecture. This enhanced 3D CNN  
incorporates a spatiotemporal attention layer designed to filter out irrelevant input sections, allowing the model  
to concentrate on critical regions. The feature extraction module uses Gray Level Co-occurrence Matrix  
(GLCM) and Local Binary Patterns (LBP) to extract texture features, which are then combined with deep  
learning features using a Single-Layer Perceptron (SLP) in the feature fusion module. Jafar et al. [17] develop  
the DFT-MF model, a deep learning approach using mouth features to detect Deepfake videos by analysing  
lip/mouth movement. The methodology involves utilising CNNs to process video frames, focusing on mouth  
movements as key indicators of manipulation. The model analyses these features to classify videos as either  
fake or real. This approach capitalises on the inconsistencies and artifacts introduced during the deepfake  
creation process, particularly in the lip and mouth areas.  
Page 3276  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
2.2 Hybrid CNN-RNN  
This model integrates a Convolutional Neural Network (CNN) with a Recurrent Neural Network (RNN),  
specifically combining Efficient Net as the CNN and Long Short-Term Memory (LSTM) as the RNN. The  
RNN layer, LSTM, is placed on top of the CNN to process sequential data by taking feature vectors extracted  
by the CNN as input. Efficient Net, pre-trained on the ImageNet dataset, is used to extract spatial features from  
input video sequences through its convolutional layers. These features are then normalised to enhance  
processing speed, activated to introduce non-linearity, and downsized using max pooling to retain essential  
information while reducing data dimensions [18]. Suratkar and Kazi [19] introduced a novel framework for  
detecting deep fake videos using transfer learning with autoencoders and a hybrid Convolutional Neural  
Network (CNN) and Recurrent Neural Network (RNN) model. The study analyses the generalisability of  
convolutional autoencoders and CNN-RNN models on benchmark datasets such as DFDC, Face Forensics,  
Face-Forensics++, and DFD. Petmezas et al., [20] presented a hybrid model with deep learning framework that  
combines CNN and Long short-term memory (LSTM) and transformer architecture to detect video based  
deepfakes. The CNN extracts the spatial facial features from the individual frames, while the LSTM captures  
temporal relationships between frames to recognize motion inconsistencies. The transformer module enhances  
attention to subtle feature and variation, improving detection accuracy and robustness.  
Sharma et al. [21] introduces a GAN-CNN ensemble model designed to improve deepfake detection in social  
media images while minimising the catastrophic forgetting issue common in CNN-based detectors. The  
methodology involves using a Generative Adversarial Network (GAN) to generate and store samples from  
previous tasks, which are then replayed during the training of new tasks. This approach aims to make the CNN  
more robust and adaptive in identifying deepfakes in dynamic environments.  
MATERIALS AND METHODS  
3.1 Data collection  
The data for this research was downloaded from Celeb-DF (v2) dataset. The integration of the Celeb-DF  
datasets provides a robust foundation for developing advanced deepfakes forensic analysis, prioritising high-  
quality synthetic videos that minimise detectable artifacts [22]. Celeb-DF contains high-quality deepfake  
videos of celebrities, generated using an improved synthesis process that reduces visual artifacts like colour  
mismatches, splicing boundaries and inconsistent face orientations. It includes 590 authentic videos of  
celebrity, additional 300 youtube real video and 5639 fake video [23]. Celeb-DF addressed limitations of  
earlier datasets (e.g., DFD, DFDC-P, FF++) by producing synthetic videos that closely resemble real-world  
deepfakes. Individual frames were extracted from the video for further processing. Video frames were  
generated by breaking down a video into individual images [2].  
3.2 Pre-processing steps  
To enhance the effectiveness of the forensic analysis of deepfakes, it is essential to first improve the image  
quality through pre-processing. A critical step involves face detection, which allows filtering out frames or  
parts of frames without faces [24]. To isolate the face Region of Interest (RoI), computer vision techniques  
were used to automatically detect faces in images. The Dlib classifier was employed to identify 68 facial  
landmarks. This detector ensembled regression trees to predict the placement of facial features based on pixel  
intensities within the images. This effectively identified faces by mapping 68 key points by eliminating  
unnecessary frames.  
The image were resized. This involves adjusting the dimensions of an image to either enlarge or reduce its size  
without altering its content. In this approach, the image resolution was standardised to 224 × 224 pixels to  
ensure consistency across the dataset. The dataset was divided into three subsets: training, validation, and  
testing. The validation phase is used to determine the most effective architecture. During model training, the  
validation set helps select the best-performing architecture, while the combined training and test sets are used  
to evaluate the model's performance after training.  
Page 3277  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
3.3 Model Development, Training and Testing  
A hybrid model that consists of 2D CNN, 1D CNN and RNN for the forensic analysis of cross-platform  
deepfakes was developed. Pre-processed video data was fed into this model, which was then trained to learn  
patterns from the input. Following training, the model was tested to evaluate its performance.  
The model employs a deep learning approach with a hybrid 2D CNN, 1D CNN and RNN to determine whether  
a video is real or fake. This classification depends on whether the number of manipulated frames in the video  
exceeds a predefined threshold. The model analyses some key variables to make this decision such as patterns,  
texture, artifacts, and temporal dynamics of consecutives frames (e.g. unnatural blinking, facial expression  
temporal incoherence). Integration of these metrics helps the model identifies inconsistencies typical of  
synthetic media, ensuring robust detection of deepfakes. The eye shape detector uses coordinates (points 37–  
46) from facial landmarks to calculate the Euclidean distances between the endpoints of both the left and right  
eyes, denoted as d1 and d2, respectively. Similarly, the lip shape detector uses lip coordinates (points 49–68) to  
compute the length of the inner lips (d3) and the outer lips (d4) by calculating the Euclidean distances among  
relevant lip coordinates. The nose shape detector utilises facial landmark data (points 28–36) to determine the  
nose's shape. It calculates the base width of the nose (d5) and the width at its highest point (d6) using  
Euclidean distances. These measurements - eye widths (d1, d2), inner and outer lip lengths (d3, d4), and nose  
widths (d5, d6) was used as shape features to train a classifier that can differentiate between real and  
manipulated videos based on variations in facial features across frames.  
3.4 Model Performance Evaluation Metrics  
This study enhanced the AI-based model performance by leveraging relevant metrics. By integrating these  
approaches, the researcher developed robust frameworks for improving AI models across diverse applications,  
ensuring higher-quality outputs and adaptability in real-world scenarios. Below are the performance metrics:  
3.5 How the model works  
The study developed a hybrid model which consists of CNN (2D and 1D), RNN for the forensic analysis of  
cross platform deep fakes, with each contributing to the model’s efficacy and efficiency. CNN models detect  
deepfakes by extracting hierarchical features to spot subtle artifacts and inconsistences in deepfakes [25]. CNN  
focuses on different facial region, while RNNs analysed the temporal sequence of video frames to detect subtle  
inconsistencies introduced in manipulated videos. 2D CNNs capture spatial features in individual frames for  
deepfakes detection, especially by analysing individual frames or images. A 2D CNN processes spatial features  
by applying convolutional filters over the two-dimensional pixel data of each frame. This extracts patterns that  
distinguish real from fake images. It is often integrated into larger video detection frameworks for better  
performance. The 1D CNN identifies local and short range pattern such as blinking anomalies, lip-sync errors  
and flickering.  
This hybrid model leverages CNN strength in spatial analysis and RNN strength in temporal sequence  
modeling. RNNs improve the accuracy of deepfakes detection by learning temporal correlations, but for its  
limitation in the detection of spatial features, the 2D CNNs was integrated to analyse single-frame spatial  
features, which cannot capture temporal inconsistencies across video frames. This are often useful cues in  
deepfake videos. The 1D CNN identify inconsistences in short sequence making it easier and faster for RNN  
by reducing the sequence length through temporal pooling.  
In the hybrid model, preprocessed video frames were first imputed into 2D CNN where individual frames were  
examined to extract spatial features. Each frame is then compress into feature vector. The vector is then fed to  
a 1D CNN that extracts temporal features over a short window by applying convolutional filters. This  
convolutional filters detect short-range and local inconsistences. The temporal pattern from 1D CNN is then  
analysed by RNN for long-range dependencies across the entire video. The result is passed through a classifier,  
where the output is classified as “real or fake”. This combination is effective because deepfakes often  
introduce temporal artifacts that static frame-based models might miss.  
Page 3278  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
Figure 1: Architecture of the model 4  
RESULTS AND DISCUSSION  
The model was developed using Jupyter Lab and machine learning libraries such as Numpy, Pandas,  
Tensorflow, Matplotlib, Hashlib, Sklearn and Seaborn for the preprocessing and training of the hybrid model.  
The choice of JupyterLab is predicated on its ability to provide an integrated platform combining code  
execution, efficient data handling, data visualisation, cleaning and documentation. The process facilitates  
flexible model development and experimentation. The collective libraries support the entire training process in  
Jupyter Lab from data manipulation it support the model construction, data visualisation and evaluation.  
The model was thoroughly trained to obtain the balance between performance and computational efficiency.  
Training and validation of the model were conducted systematically using data subset obtained from Celeb DF  
with the training taking placed on the selected training set and valuation of the set for validation; this is to help  
prevent overfitting and accurate assessment of the model. The model was trained for 100 epoch to monitor  
peak of the model performance.  
Figure 2: Training and Validation AUC  
The model AUC climbs steadily from 0.90 at epoch 1 and reaches its peak at 0.99 for epoch 100. This implies  
that the model fits better with the training leading to the standardisation of the model at 100 epoch. Figure 2  
describes the model AUC with the blue line representing the training AUC sets and the yellow line indicating  
Page 3279  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
the validation AUC. The increasing AUC value indicate an excellent discriminative power, which is the ability  
of the model to distinguish between real and fake videos.  
Figure 3: Training and Validation Loss  
Figure 3 shows the loss in the training and the validation sets. The result revealed that the value loss of training  
set at 100 epoch was 0.05, while that of the validation set was 0.18. This indicates that there is no overfitting in  
model. The value loss indicates how well the training and validation datasets were predicted by the model.  
Lower value loss is better. The model training loss decrease generally and the training accuracy increases,  
meaning the model is learning from the data.  
Figure 4: Training and Validation Precision/Recall  
Precision measures the accuracy of positive prediction while Recall measures how well the model captures  
actual positives. Figure 4 shows steady improvement over 100 epochs for both training and validation sets.  
Training precision and recall reached about 0.98 while validation precision and recall stabilised around 0.95.  
The close range between training and validation metrics indicates good generalisation and balanced  
performance, with the model effectively managing false positive and false negatives. Overall, the model  
Page 3280  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
demonstrates strong and reliable learning with well-balanced precision and recall reflecting a successful  
training process and architecture.  
Figure 5: Training and Validation Accuracy  
Accuracy is a common metric used to evaluate the performance of a classification model. It measures the  
proportion of correctly predicted instances out of the total instances evaluated. Figure 5 shows improvement in  
both training and validation accuracy across 100 epochs. Initially, the model starts with about 90% accuracy on  
both training and validation data, indicating a reasonable baseline performance from the early stages. As the  
epochs progress, the training accuracy rises smoothly, eventually reaching approximately 99%. This  
continuous increase reflects the model’s growing ability to learn and fit the training data by effectively  
capturing the relevant features and patterns. The validation accuracy also improves steadily, increasing from  
around 90% to nearly 99% by the end of 100 epochs. The close tracking of validation accuracy alongside  
training accuracy suggests the model generalises well to unseen data and does not merely memorise the  
training set. The relatively small gap between training and validation accuracy indicates limited overfitting  
which is a desirable traits because it shows the model predictions are robust and reliable when applied to new  
examples.  
4.1 Model testing  
The model was tested on a balanced dataset of 200 videos, consisting of 100 real and 100 fake videos to  
evaluate the performance of the hybrid deepfake video detection model. The objective was to assess the  
model’s effectiveness in accurately classifying videos as either fake or real and this objective was achieved as  
the model was able to identify video of fraudulent origin.  
Figure 6 shows the confusion matrix of the text data. The result falls into both fake and real.  
Fig 6: Confusion matrix for test data  
The results below show what was found after testing 200 videos of fake and real videos:  
Page 3281  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
True Positive (TP) is the number of cases that the model was able to predict the positive class in respect to the  
ground truth “Fake” therefore TP are videos that are actually predicted by the model as fake.  
TP = 90  
True Negative (TN): is a situation where the model predicts a negative outcome correctly. True negative in  
this context is when the model correctly predict real video as real when it is actually real.  
TN = 97  
False Positive (FP) is when the model predicts the positive outcome incorrectly. In this context, it is when the  
model predicts a real video as fake.  
FP = 3  
False Negative (FN) is when the model predicts a negative outcome incorrectly. In this case, it occurs when  
the model predicts fake video as real.  
FN = 10  
The performance of the developed model was evaluated based on different metrics. The model classified 187  
out of the 200 test data correctly, thereby achieving an overall accuracy of 93.5%  
+
=
=
= 0.935  
=
+
+
+
The model achieved a precision of 96.8%, which means that almost all the videos it identified as deepfakes  
were truly manipulated. However, there were a few false positives. Similarly, the model reached a recall of  
90%, which shows that most of the deepfake videos were correctly detected, with only 10% incorrectly  
classified as real.  
=
=
= 0.90  
=
(2)  
+
90  
90  
== = 0.968  
93  
=
(3)  
+
90 + 3  
3
3
=
=
= = 0.03  
(4)  
+
3 + 97  
100  
The F1 Score which is 93.3% shows the balance between precision and recall, this is a proof that the classifier  
is reliable and can effectively handle false positives and false negatives.  
2 × (  
×
)
2 × (0.968 × 0.9)  
1
=
== 0.933  
(5)  
+
0.968 + 0.90  
With a Specificity of 97%, the model has shown its ability to accurately differentiate between real and fake  
videos. The balanced accuracy of 93.5% makes it clear that the model’s performance is consistently good  
across both classes. Hence, ensuring there is no partiality between real and fake videos.  
Page 3282  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
=
=
= 0.97  
=
(6)  
+
(
)+  
=
=
(7) 2  
0
. = 0.935  
(8)  
The Matthews Correlation Coefficient (MCC) helps to get a balanced evaluation. The MCC of 0.872 confirms  
that the model is reliable. Since the MCC is close to 1, it simply means the model performed greatly in  
predicting real and fake videos  
90 97  
×
3
10  
)−( × )  
(
×
)−(  
×
)
(
(10)  
Table 4.1 is a summary of the metrics used to evaluate the performance of the model test data and their  
respective values?  
Figure 7 shows the ROC curve. AUC is derived from the ROC, and it is an essential metrics. AUC is an  
important statistical metrics utilise due to the imbalance of the dataset. AUC denotes the area under the curve  
of a plot of False Positive Rate versus True Positive Rate at various locations in the interval [0, 1]. As the value  
increases the model's performance improves. This model's AUC score is 0.935, indicating that it has a 93.5%  
probability of successfully identifying a fake video.  
Figure 7: Receiver Operating Characteristic Table 1 Summary of the Model’s Performance Metrics  
Table 1 Summary of the Model’s Performance Metrics  
S/No  
Metric  
Value  
1
2
3
Accuracy  
Precision  
93.5%  
96.8%  
90.0%  
Recall  
(Sensit  
ivity)  
4
Specificity  
97.0%  
Page 3283  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
5
F1 Score  
93.3%  
6
7
Balanced Accuracy  
MCC  
93.5%  
0.872  
8
AUC  
0.935  
DISCUSSION OF THE FINDINGS  
The hybrid model was developed with the intention of capturing spatial and temporal features in deepfakes.  
Deepfakes often leave a trace of artifacts such as inconsistences in colour, texture, and irregular edges. There  
are also inconsistences in facial movement and unnatural blinking pattern [26,4]. Considering the training and  
validation performance as indicated from result of metrics, it can be deduced that the model learned the spatial  
and temporal features effectively from the dataset (Celeb DF) and there is no evidence of overfitting. The near  
perfect training AUC and total accuracy reflect the strong classification power of the model to distinguish  
between real video and deepfakes. The training precision of 98% indicates that the model is reliable in  
predicting true positive. According to Cuellar [27], precision is a very important forensic metric where false  
positives undermine confidence score and lead to waste in the investigation resources. The model training and  
validation precision demonstrate a high ability to minimise false alarm. The high training and validation recall  
rate shows the ability of the model to detect manipulation in deepfakes while maintaining a high precision. The  
architectural design, the training and validation metrics align with superiority of hybrid models over CNN or  
RNN single model. The result of the various metrics shows the superiority of the CNN-2D-RNN hybrid model  
over individual model and similar hybrid model [20].  
The hybrid model achieved an accuracy of 93.5%, precision of 96.8%, recall of 90%, F1 score of 93.3%, MCC  
of 0.872 and AUC of 0.935 on the test dataset obtained from Celeb DF. These results show that the model  
performed better than similar CNN, RNN and a CNN-RNN hybrid model reported [20]. The AUC and  
accuracy value on the test data (Celeb DF) reflect the strong classification power of the model to distinguish  
between real video and deepfakes. The precision of 96.8% on the test data indicate that the model is reliable in  
predicting true positive, precision is a very important forensic metric where false positives undermine  
confidence score and lead to waste in the investigation resources. The model testing precision demonstrates a  
high ability to minimise false alarm as the false positive rate is only 3.2%. The recall rate shows the ability of  
the model to detect manipulation in deepfakes while maintaining a high precision. Matthew Correlation  
Coefficient (MCC) has not been given much attention in the past, it is a strong way to measure binary  
classification because it incorporates all confusion matrix [12]. It has been argued to be better than F1 and  
accuracy in terms of imbalance. The F1 score of 93.3% shows a balance between precision and recall. In some  
cases, high recalls are achieved at the expense of precision and vice versa. The reported MCC of 0.872 shows  
that the developed model is better at balancing true and false predictions. The use of MCC metric makes the  
model more reliable than many other methods that have not reported. The model specificity is 97% meaning  
the correctly classifying 97% of real video. This reduces the risk of labelling real video as fake. Unlike some  
models, this model maintains a balance between specificity and recall.  
The current results show that the developed hybrid CNN / 2D-CNN / RNN model not only outperforms  
previous CNN-based and hybrid models in accuracy and precision but also offers a stronger basis for real-  
world deployment in forensic and security-sensitive environments.  
5.1  
Limitation  
This research encountered a few limitations. The research was carried out on a limited dataset using only data  
from Celeb DF for training validation and testing. The high cost of hardware and lack of funding were critical  
challenges in carrying out this research.  
Page 3284  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
CONCLUSION  
In this research work, we developed, validated and tested an AI-based hybrid model consisting of 2D CNN, 1D  
CNN and RNN architecture for a reliable and improved cross-platform forensic analysis of deepfakes with  
high accuracy. The application of CNN component of the hybrid extracts the spatial features frame by frame  
while the RNN component captures the temporal features, the model was able to capture effectively the  
temporal inconsistency and subtle artifacts introduced during the synthesis of the deepfakes. The hybrid model  
performed highly in metrics like accuracy 93.5 %, Precision 96.8% Recall 90 % and MCC 87.2 %. The  
application of many metrics particularly the MCC demonstrates the balanced and unbiased nature of the hybrid  
model and this presents the quality and reliability of the model for forensic admissibility. This model  
outperformed individual CNN and RNN models in adaptability and generality in addressing emerging  
deepfakes technology. Worthy of note in this research is finding on the importance of involving hybrid  
architecture in digital forensics for robust and trustworthy results. It can be deduced that this study established  
a solid foundation for employing hybrid model for digital forensics in solving the security posed by deepfakes.  
Future study will focus on testing this hybrid model on other databases and comparing it with other  
transformer based models.  
Funding: No funding was received for this research, as the research was self-funded.  
Conflict of Interest: There is no conflict of interest.  
Dual Publication: No dual publication submission.  
Ethical Approval: Not applicable  
REFERENCES  
1. Agarwal, H., & Singh, A. (2021). Deepfake detection using svm. In 2021 Second International  
Conference on Electronics and Sustainable Communication Systems (ICESC) 1245-1249.  
2. Reis, P. M. G. I., & Ribeiro, R. O. (2024). A forensic evaluation method for DeepFake detection  
using DCNN-based facial similarity scores. Forensic Science International, 358, 111747.  
3. Chen, H. S., Rouhsedaghat, M., Ghani, H., Hu, S., You, S., & Kuo, C. C. J. (2021). Defakehop: A  
lightweight high-performance deepfake detector. In 2021 IEEE International conference on  
Multimedia and Expo (ICME) 1-6.  
4. Sunil, R., Mer, P., Diwan, A., Mahadeva, R., & Sharma, A. (2025). Exploring autonomous methods  
for deepfake detection: A detailed survey on techniques and evaluation. Heliyon e42273.  
5. Choi, Y., Uh, Y., Yoo, J., & Ha, J. W. (2020). Stargan v2: Diverse image synthesis for multiple  
domains.  
In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp.  
81888197).  
6. Chen, H., & Magramo, K. (2024). Finance worker pays out $25 million after video call with  
deepfake ‘chief financial officer’(2024). URL https://www. cnn. com/2024/02/04/asia/deepfake-cfo-  
scam-hong-kong-intlhnk/index. html.  
7. Arshad, H., Jantan, A. B., & Abiodun, O. I. (2018). Digital forensics: Review of issues in scientific  
validation of digital evidence. Journal of Information Processing Systems, 14(2).  
8. Sharma, B., Ghawaly, J., McCleary, K., Webb, A. M., & Baggili, I. (2025). ForensicLLM: A local  
large language model for digital forensics. Forensic Science International: Digital Investigation, 52,  
301872.  
9. Lukas, N., Salem, A., Sim, R., Tople, S., Wutschitz, L., & Zanella-Béguelin, S. (2023). Analysing  
leakage of personally identifiable information in language models. In 2023 IEEE Symposium on  
Security and Privacy (SP) 346-363.  
10. Horsman, G., & Dodd, A. (2024). Competence in digital forensics. Forensic Science International:  
Digital Investigation, 51, 301840.  
Page 3285  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
11. de Rancourt-Raymond, A., & Smaili, N. (2023). The unethical use of deepfakes. Journal of Financial  
Crime, 30(4), 1066-1077.  
12. Chicco, D., & Jurman, G. (2023). The Matthews correlation coefficient (MCC) should replace the  
ROC AUC as the standard metric for assessing binary classification. BioData Mining, 16(1), 4.  
13. Viola, M., & Voto, C. (2023). Designed to abuse? Deepfakes and the non-consensual diffusion of  
intimate images. Synthese, 201(1), 30.  
14. Batagelj, B., Kronovšek, A., Štruc, V., & Peer, P. (2025). Robust cross-dataset deepfake detection  
with multitask self-supervised learning. ICT Express.  
15. Kosarkar, U., Sarkarkar, G., & Gedam, S. (2023). Revealing and classification of deepfakes video's  
images using a customize convolution neural network model. Procedia Computer Science, 218,  
2636-2652.  
16. Almestekawy, A., Zayed, H. H., & Taha, A. (2024). Deepfake detection: Enhancing performance  
with spatiotemporal texture and deep learning feature fusion. Egyptian Informatics Journal, 27,  
100535.  
17. Jafar, M. T., Ababneh, M., Al-Zoube, M., & Elhassan, A. (2020). Forensics and analysis of deepfake  
videos. In 2020 11th international conference on information and communication systems (ICICS)  
053058.  
18. Koritala, S. P., Chimata, M., Polavarapu, S. N., Vangapandu, B. S., Gogineni, T. K., & Manikandan,  
V. M. (2024, June). A Deepfake detection technique using Recurrent Neural Network and  
EfficientNet. In 2024 15th International Conference on Computing Communication and Networking  
Technologies (ICCCNT) (pp. 1-6). IEEE.  
19. Suratkar, S., & Kazi, F. (2023). Deep fake video detection using transfer learning approach. Arabian  
Journal for Science and Engineering, 48(8), 9727-9737.  
20. Petmezas, G., Vanian, V., Konstantoudakis, K., Almaloglou, E. E., & Zarpalas, D. (2025). Video  
deepfake detection using a hybrid CNN-LSTM-Transformer model for identity verification.  
Multimedia Tools and Applications, 1-20.  
21. Sharma, P., Kumar, M., & Sharma, H. K. (2024). GAN-CNN ensemble: a robust deepfake detection  
model of social media images using minimised catastrophic forgetting and generative replay  
technique. Procedia Computer Science, 235, 948-960.  
22. Li, Y., Yang, X., Sun, P., Qi, H., & Lyu, S. (2020). Celeb-df: A large-scale challenging dataset for  
deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern  
recognition (pp. 3207-3216).  
23. Vamsi, V. V. V. N. S., Shet, S. S., Reddy, S. S. M., Rose, S. S., Shetty, S. R., Sathvika, S., ... &  
Shankar, S. P. (2022). Deepfake detection in digital media forensics. Global Transitions Proceedings,  
3(1), 74-79.  
24. Deng, J., Guo, J., Ververas, E., Kotsia, I., & Zafeiriou, S. (2020). Retinaface: Single-shot multi-level  
face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and  
pattern recognition (pp. 5203-5212).  
25. Dong, F., Zou, X., Wang, J., & Liu, X. (2023). Contrastive learning-based general Deepfake  
detection with multi-scale RGB frequency clues. Journal of King Saud UniversityComputer and  
Information Sciences, 35(4), 90-99.  
26. Sohail, S., Sajjad, S. M., Zafar, A., Iqbal, Z., Muhammad, Z., & Kazim, M. (2025). Deepfake Image  
Forensics for Privacy Protection and Authenticity Using Deep Learning. Information, 16(4), 270.  
27. Cuellar, M. (2024). The Neglected Error: False Negatives and the Case for Validating Eliminations.  
arXiv preprint arXiv:2412.05398.  
Page 3286  
www.rsisinternational.org