0% found this document useful (0 votes)
15 views14 pages

Deepfake Detection Using Hybrid Models

The document presents a synopsis of a project aimed at developing a hybrid deepfake detection system using Convolutional Neural Networks (CNN), ResNeXt, and Long Short-Term Memory (LSTM) networks to enhance detection accuracy of manipulated digital content. The proposed model integrates spatial and temporal analysis to effectively identify deepfakes, addressing the growing concerns of misinformation and digital security. The project outlines a comprehensive methodology, including data collection, preprocessing, model design, training, and evaluation, with an expected outcome of achieving over 90% accuracy in real-time detection.

Uploaded by

akash96243
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
15 views14 pages

Deepfake Detection Using Hybrid Models

The document presents a synopsis of a project aimed at developing a hybrid deepfake detection system using Convolutional Neural Networks (CNN), ResNeXt, and Long Short-Term Memory (LSTM) networks to enhance detection accuracy of manipulated digital content. The proposed model integrates spatial and temporal analysis to effectively identify deepfakes, addressing the growing concerns of misinformation and digital security. The project outlines a comprehensive methodology, including data collection, preprocessing, model design, training, and evaluation, with an expected outcome of achieving over 90% accuracy in real-time detection.

Uploaded by

akash96243
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

A

SYNOPSIS

ON

Deep Fake Detection System

Submitted in partial fulfillment of the requirements

for the award of the degree of

Bachelor of Technology
in

Computer Science and Engineering

By

Sahil Akash, Tanuj Kumar, Sachin Patel

(2300970100163, 2300970100204, 2300970100162)

Semester – V

Under the Supervision of


Dr. Lopamudra Mohanty

Galgotias College of Engineering & Technology

Greater Noida 201306

Affiliated to

Dr. APJ Abdul Kalam Technical University, Lucknow

October, 2025

2
Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Literature Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Familiarization with the problem . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Work Done by Others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Research Methodology Insights . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

4 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

5 Methodology / Planning of Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 9


5.1 Phase 1: Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.2 Phase 2: Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.3 Phase 3: Model Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.4 Phase 4: Training & Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.5 Phase 5: Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.6 Phase 6: Testing & Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.7 Tools and Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.8 Expected Outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

6 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1
Abstract
With the rapid growth of artificial intelligence, deepfakes—realistic fake images and videos
created using AI—have become a major concern for online safety and digital trust. Detecting
these manipulations is important to prevent misinformation and misuse of digital content.
In this project, we propose a hybrid deepfake detection model that combines the strengths
of three approaches: Convolutional Neural Networks (CNN), ResNeXt, and Long Short-Term
Memory (LSTM) networks. CNN and ResNeXt are used to capture visual and spatial details
from images, while LSTM helps in learning the sequence or temporal patterns from video
frames.
Each model is trained separately on real and fake facial data, and the final decision is made
by averaging their individual accuracies or prediction outputs. The results show that the hybrid
model performs better than using any single model alone. This study aims to support the
development of more dependable systems that can identify fake digital media and help maintain
online authenticity.

Keywords: Deepfake Detection, Hybrid Model, CNN, ResNeXt, LSTM, Ensemble Learning,
Digital Media Authenticity.

2
1 Introduction
In today’s digital world, multimedia content such as videos and images play a vital role in com-
munication, entertainment, and information sharing. However, the advancement of Artificial
Intelligence (AI) and deep learning has also given rise to a concerning technology known as
deepfakes. Deepfakes are synthetic media where a person’s likeness is replaced or altered us-
ing machine learning algorithms, often making it difficult to distinguish between real and fake
content.
This has led to growing threats such as misinformation, identity misuse, cyberbullying, and
damage to reputation. Therefore, the ability to detect and prevent deepfakes has become an
important research area in both academic and professional domains.

Figure 1: Comparison of Real and Deepfake Face Images

Early detection methods mainly relied on handcrafted features such as facial landmarks
or inconsistencies in lighting. However, these traditional approaches often fail when deepfakes
become more sophisticated. Recent advances in deep learning have shown that neural networks
can automatically extract complex patterns that help distinguish fake content from authentic
ones. Yet, using a single model often limits performance due to data variability and different
types of manipulations.
To overcome these challenges, this project proposes a hybrid model that integrates mul-
tiple architectures—Convolutional Neural Network (CNN), ResNeXt, and Long Short-Term
Memory (LSTM)—to improve detection accuracy.

• CNN captures low-level spatial features such as textures and edges.

• ResNeXt, a deeper variant of ResNet, focuses on learning more abstract and robust
visual features.

• LSTM models the temporal or sequential relationships in video frames, which


helps in identifying subtle motion irregularities present in fake videos.

3
By combining these models, the system takes advantage of both spatial and temporal cues,
making the overall detection more reliable. Each model is trained on real and fake facial image
datasets, and their predictions are averaged to produce a final output. This ensemble approach
increases overall performance and reduces the error caused by relying on a single model.

Figure 2: Architecture of the Proposed Hybrid Detection Model

The proposed approach contributes toward developing a more robust and practical solution
for real-world deepfake detection. It also encourages the exploration of hybrid architectures
that blend the power of different neural networks.
Through this project, we aim not only to achieve accurate detection but also to raise aware-
ness about the ethical and security challenges associated with deepfake technology.

4
2 Literature Survey

2.1 Familiarization with the problem

Deepfake technology uses advanced deep learning models, especially Generative Adversarial
Networks (GANs), to create highly realistic fake videos and images that mimic real people.
While it has positive applications in film and media, its misuse for spreading misinformation,
identity theft, and fraud poses a serious threat to digital security. Traditional detection methods
based on visible artifacts are no longer effective as deepfakes have become increasingly sophis-
ticated. Therefore, there is a growing need for intelligent systems capable of identifying subtle
inconsistencies. Hybrid deep learning approaches combining CNN, ResNeXt, and LSTM mod-
els are emerging as promising solutions, as they can analyze both spatial and temporal features
to detect manipulated content more accurately and reliably.

2.2 Work Done by Others

• Jaiswal et al. (2022) proposed a hybrid CNN-LSTM model for detecting deepfakes
in short video clips using Indian datasets from YouTube and DFDC.

• Kumar and Sinha (2023) implemented a ResNeXt-based architecture trained on


synthetic datasets to detect manipulated facial regions, improving feature extraction
efficiency.

• Patel et al. (2024) investigated the use of transfer learning with pretrained CNNs
(VGG16, InceptionV3) for low-resource environments, addressing computational
limitations in local systems.

• Li et al. (2020) proposed “Face X-Ray,” a CNN-based approach to detect blending


artifacts left by face-swapping algorithms. The method achieved high performance
in identifying manipulated faces but struggled with low-quality videos.

• Afchar et al. (2018) introduced MesoNet, a lightweight CNN architecture de-


signed for efficient real-time deepfake detection. While efficient, it had limited
robustness to unseen datasets.

5
• Nguyen et al. (2019) developed a Capsule Network-based model to detect deep-
fakes by capturing spatial relationships between facial features, achieving 96% ac-
curacy on benchmark datasets.

• Korshunov and Marcel (2019) analyzed the impact of compression on deepfake


detection models and found that performance drops significantly under high com-
pression ratios, emphasizing preprocessing importance.

• Dolhansky et al. (2020) released the Deepfake Detection Challenge (DFDC) dataset,
enabling large-scale benchmarking for deepfake detection models and motivating
hybrid ML approaches.

2.3 Research Methodology Insights

The hybrid CNN–ResNeXt–LSTM model combines:

1. CNN for low-level spatial feature extraction.

2. ResNeXt for deep hierarchical and residual features.

3. LSTM for sequential frame-based temporal analysis.

4. Averaging or weighted accuracy fusion can provide a more stable and generalized
output.

The review suggests structuring methodology with:

1. Multi-dataset training (e.g., DFDC, FaceForensics++).

2. Transfer learning for initialization.

3. Cross-validation to reduce overfitting.

6
3 Problem Formulation
The increasing use of AI for generating synthetic media has created major social and cyberse-
curity threats. Deepfakes can manipulate perceptions, spread misinformation, and damage rep-
utations. Existing detection systems often fail when tested on unseen datasets or compressed
videos.
Therefore, there is a pressing need for a robust, scalable, and adaptive detection model ca-
pable of analyzing both images and videos. The proposed system addresses these challenges
by combining CNN-based visual learning with temporal sequence analysis through LSTM net-
works, creating a comprehensive detection framework that can handle diverse deepfake manip-
ulation techniques.
The significance of this work lies in its potential to:

• Protect individuals and organizations from identity theft and reputational damage

• Combat the spread of misinformation in digital media

• Provide a scalable solution for real-time deepfake detection

• Contribute to the broader field of digital media forensics and authentication

7
4 Objectives
1. Integrate features from both lightweight (MobileNetV2) and deep (ResNet50) CNNs
into a hybrid detection model to combine efficiency and accuracy in identifying
deepfake artifacts.

2. Utilize pretrained CNN architectures such as MobileNetV2 and ResNet50 to extract


robust spatial features from video frames, reducing training time and improving
model performance on limited datasets.

3. Train and evaluate the model on benchmark datasets including Celeb-DF and DFDC
to ensure comprehensive performance assessment.

4. Prioritize model efficiency by using MobileNetV2 for faster inference, making the
hybrid model suitable for real-time or low-resource deepfake detection scenarios
without compromising accuracy.

5. Analyze model accuracy, precision, recall, and F1-score for comprehensive perfor-
mance evaluation across different types of deepfake manipulation techniques.

8
5 Methodology / Planning of Work
The project follows an iterative machine learning workflow divided into multiple phases. Figure
3 illustrates the complete system workflow from input to output.
The methodology is structured as follows:

5.1 Phase 1: Data Collection

Gather benchmark datasets including Celeb-DF and DFDC (Deepfake Detection Challenge)
containing both real and deepfake videos. These datasets provide diverse manipulation tech-
niques and quality levels for robust model training.

5.2 Phase 2: Preprocessing

• Extract individual frames from video sequences at consistent intervals

• Detect and align faces using MTCNN (Multi-task Cascaded Convolutional Net-
works)

• Resize and normalize frames to match CNN input requirements

• Apply data augmentation techniques to improve model generalization

5.3 Phase 3: Model Design

Build a hybrid architecture incorporating:

• CNN layers for spatial feature extraction

• ResNeXt architecture for deep visual representation learning

• LSTM networks for temporal sequence analysis

• Ensemble mechanism for combining predictions from multiple models

5.4 Phase 4: Training & Evaluation

Train individual models on preprocessed datasets, optimize hyperparameters using validation


sets, and evaluate performance using standard metrics including accuracy, precision, recall, and
F1-score.

9
Figure 3: Complete System Workflow of Deep Fake Detection System

10
5.5 Phase 5: Integration

Combine frame-wise predictions from individual models using ensemble averaging or voting
mechanisms to determine the authenticity of entire video sequences.

5.6 Phase 6: Testing & Validation

Test the integrated system on unseen data and perform robustness analysis under various con-
ditions including compression, resolution changes, and different manipulation techniques.

5.7 Tools and Technologies

Backend Development:

• Programming Language: Python 3.8+

• Deep Learning Frameworks: TensorFlow/Keras, PyTorch

• Computer Vision Libraries: OpenCV, MTCNN

• Data Processing: NumPy, Pandas

• Web Framework: Flask/FastAPI for REST API development

• Model Serving: TensorFlow Serving or PyTorch Serve

• Database: MongoDB/PostgreSQL for storing detection results

Frontend Development:

Programming Language: HTML,CSS,Javascript

5.8 Expected Outcome

A hybrid detection system capable of identifying deepfakes in real-time with over 90% accu-
racy on benchmark datasets, demonstrating robust performance across various manipulation
techniques and video quality conditions.

11
6 References
[1] Khan, M. A., Artusi, A., & Dai, Q. “Adversarially Robust Deepfake Media Detection
Using Fused Convolutional Neural Network Predictions,” arXiv preprint, arXiv:2102.05950,
2021.

[2] Saikia, S., et al. “A Hybrid CNN-LSTM Model for Video Deepfake Detection by
Leveraging Optical Flow Features,” arXiv preprint, arXiv:2208.00788, 2022.

[3] Nadimpalli, M., & Rattani, A. “On Improving Cross-dataset Generalization of Deep-
fake Detectors,” arXiv preprint, arXiv:2204.04285, 2022.

[4] Angeline, K. D., & Kusniyati, M. “Komparasi Performa VGG19, ResNet50, DenseNet121
dan MobileNetV2 dalam Mendeteksi Gambar Deepfake,” Jurnal CESS, vol. 8, no. 1,
pp. 317–324, 2023.

[5] Ikram, S., et al. “A Performance Enhancement of Deepfake Video Detection through
the Use of a Hybrid CNN Deep Learning Model,” International Journal of Electrical
and Computer Engineering Systems, vol. 14, no. 1, pp. 43–50, 2023.

[6] Al-Adwan, A., et al. “Detection of Deepfake Media Using a Hybrid CNN–RNN
Model and Particle Swarm Optimization (PSO) Algorithm,” Computers, vol. 13, no.
4, 2024.

[7] Maguluri, A., et al. “Hybrid Deepfake Detection Using CNN for Spatial Analysis and
LSTM for Temporal Consistency,” IJISRT, vol. 10, no. 3, 2025.

12

Common questions

Powered by AI

Sophisticated deepfake creation techniques using GANs (Generative Adversarial Networks) produce highly realistic media that traditional detection methods, which often rely on visible artifacts or manual feature extraction, struggle to identify . The innovation introduced by the hybrid model to overcome these challenges includes the amalgamation of CNN, ResNeXt, and LSTM architectures. This integration enhances the system’s ability to automatically extract complex spatial and temporal features, providing a comprehensive analysis that significantly improves detection accuracy, even in the face of subtle manipulations and high-quality fakes that traditional methods fail to capture .

The expected outcomes of implementing the hybrid deepfake detection system include achieving over 90% accuracy on benchmark datasets, demonstrating robust performance across a variety of manipulation techniques and video qualities . These outcomes align with its objectives of providing a real-time, scalable solution for detecting deepfakes, ensuring robust generalization through multi-architecture integration, and effectively addressing misinformation and identity protection challenges. The hybrid model's design and its use of advanced concept-driven architecture fulfill its goal of improving the detection efficacy against diverse deepfake techniques, which are essential for real-world applications and academic research in digital media forensics .

Transfer learning in the proposed deepfake detection system helps by initializing models with pre-trained knowledge, significantly reducing the time and computational resources needed for training on new data. It allows the detection system to leverage learned features from extensive datasets to improve model performance, particularly in low-resource environments or when dealing with limited training datasets . This approach enhances the model's ability to generalize, ensuring accurate extraction of relevant features needed for identifying deepfakes across different scenarios .

The proposed hybrid model enhances real-world application reliability for deepfake detection by combining spatial and temporal analysis capabilities, making it adept at handling diverse manipulations in multimedia content. The integration of CNN, ResNeXt, and LSTM ensures the model can capture detailed visual features and motion patterns effectively, thus improving accuracy in identifying fake content across varied quality levels. Moreover, the model's design is focused on real-time applicability, utilizing efficient architectures like MobileNetV2 to ensure fast inference suitable for practical deployment, without sacrificing accuracy . This approach makes the detection system robust to unseen data variations and different compression scenarios, crucial for real-world conditions .

Video compression can significantly reduce the effectiveness of deepfake detection models by altering the artifacts that these models rely on for identification, thus decreasing overall performance . The proposed hybrid detection system addresses this issue by employing robust preprocessing techniques that adapt to various video qualities and resolutions. It processes input data with consistent frame extraction and normalization techniques, ensuring each model component can maintain high-fidelity data analysis, thereby reducing the negative impact of compression on model accuracy .

The advantage of using a hybrid model combining CNN, ResNeXt, and LSTM lies in its ability to leverage the strengths of each of these architectures to improve deepfake detection accuracy. CNN captures low-level spatial features such as textures and edges, ResNeXt enhances the model's capacity to learn abstract and robust visual features, and LSTM models the temporal or sequential relationships in video frames, crucial for identifying motion irregularities in fake videos . Combining these models helps the detection system to effectively utilize both spatial and temporal cues, leading to more reliable results than using any single model .

The proposed hybrid model supports the goals of digital media forensics by providing a scalable, adaptive framework that accurately identifies deepfake manipulations. By combining spatial analysis through CNNs with temporal sequence detection using LSTMs, the model ensures thorough examination of media authenticity, which is critical for forensic investigations. It aids in protecting digital identities and combating misinformation, thus contributing to maintaining the integrity of digital information. Additionally, the model's ability to perform in real-time and adapt to different datasets makes it a valuable tool for forensic experts dealing with a wide range of digital media manipulation cases .

The hybrid deepfake detection model employs ensemble averaging or voting mechanisms to improve decision accuracy by combining predictions from multiple models. After each component model (CNN, ResNeXt, LSTM) generates predictions, these are aggregated either by averaging the results or through majority voting to produce a final decision on whether content is real or fake. This ensemble approach helps mitigate the limitations of individual models by leveraging their combined strengths, thereby enhancing the system's overall accuracy and robustness against various deepfake techniques .

The hybrid deepfake detection model enhances its training and evaluation effectiveness through multiple strategies: training on large-scale, diverse datasets such as DFDC and FaceForensics++ to ensure robustness across different manipulation techniques; using transfer learning with pretrained models to efficiently extract features; applying data augmentation to improve generalization; and employing cross-validation to minimize overfitting. The evaluation process includes optimizing hyperparameters and assessing performance using metrics like accuracy, precision, recall, and F1-score .

Using a single model for deepfake detection often leads to limited performance because these models might fail to generalize across different types of data manipulations and may miss certain artifacts or patterns due to their specific focus, such as spatial or temporal features . The hybrid model addresses these challenges by integrating CNN, ResNeXt, and LSTM frameworks. This integration allows the model to capture both low-level spatial details and complex temporal sequences, thereby improving detection accuracy and robustness against various manipulation techniques and unseen data variations .

You might also like