reputations and prevent them from becoming

Introduction victims of malicious attacks.

The problem addressed by this project is the
proliferation of deepfake videos and the
associated risks they pose. Deepfake 3.Preserving Trust in Media: The prevalence
technology utilizes deep learning algorithms of deepfake videos raises concerns about the
to create realistic-looking videos where trustworthiness of digital media. By
individuals appear to say or do things they developing effective detection techniques,
never actually did. These videos can be used trust in digital content can be reinforced,
maliciously to spread misinformation, ensuring that viewers can discern between
manipulate public opinion, or defame authentic and manipulated videos.
individuals. As deepfake technology becomes
increasingly sophisticated and accessible, the
need for reliable detection methods becomes 4.Safeguarding National Security:** Deepfake
paramount to mitigate the potential harm technology poses a threat to national security
caused by these videos. Therefore, this by potentially influencing public opinion,
project aims to develop an effective deep destabilizing political discourse, or inciting
learning-based model for detecting deepfake social unrest. Detecting and neutralizing
videos, thereby helping to combat the spread malicious deepfake content can help mitigate
of misinformation and protect individuals these risks and protect democratic processes.
from potential harm.

5.Advancing Technology Ethics: Addressing

the challenges posed by deepfake videos
Significance contributes to the ongoing discourse on
The importance of developing an effective ethical considerations in technology
deepfake video detection model lies in development. By promoting responsible use
safeguarding the integrity of digital content of AI and deep learning, this project aligns
and preserving trust in media sources. Here with broader efforts to ensure that
are several key reasons why this project is technological advancements benefit society
significant: as a whole.

1.Mitigating Misinformation: Deepfake videos

have the potential to deceive viewers by Overall, the development of a deepfake video
portraying false events or statements as detection model holds significant importance
genuine. Detecting and flagging such videos in maintaining the integrity of digital media,
can help prevent the spread of protecting individuals' reputations, and
misinformation, thereby preserving the upholding trust in information sources in an
accuracy of information available to the public. increasingly digital and interconnected world.

2.Protecting Individuals: Deepfake videos can

be used to manipulate and defame Summarization of the
individuals by attributing false actions or implemented system:
statements to them. A robust detection
model can help protect individuals' The project focuses on developing a deepfake
detection model using PyTorch, aiming for
unbiased training by using an equal number including facial landmarks, micro-expressions,
of real and fake videos. The dataset is sourced and inconsistencies in lighting and shadows.
from various repositories like DFDC,
FaceForensic++, and Celeb-DF, ensuring
diversity and representativeness. 1.Deep Learning Models: Deep learning
Preprocessing involves isolating facial regions models, particularly convolutional neural
from videos, standardizing them to 112 x 112 networks (CNNs) and recurrent neural
pixels at 30 fps, and splitting them into frames networks (RNNs), are commonly used for
for analysis. A threshold of 150 frames per deepfake detection. These models are trained
video is set to manage computational on large datasets to learn patterns indicative
resources. The dataset is then split into of manipulated videos.
training (70%) and testing (30%) sets,
ensuring equal representation of real and fake
videos in each subset. The model architecture 2.Multimodal Approaches: Combining
combines CNNs (ResNext) for feature information from multiple modalities, such as
extraction and an LSTM for sequential analysis visual, audio, and metadata, has shown
and classification, leveraging temporal promise in enhancing detection accuracy.
information for accurate detection. Multimodal approaches leverage
Hyperparameters are finetuned using the complementary information to improve the
Adam optimizer, with a learning rate of 1e-5 robustness of detection systems.
and weight decay of 1e-3.

3.Adversarial Techniques: Adversarial training

Implementation Survey and detection methods aim to create models
that are resilient to adversarial attacks. These
Work related to deepfake video detection techniques involve training models with
involves a range of approaches aimed at adversarial examples to improve their
identifying and mitigating the spread of robustness against manipulation attempts.
manipulated videos. Researchers have
explored various techniques to combat the
increasing sophistication of deepfake 4.Explainability and Interpretability: Ensuring
technology. Some key areas of focus include: transparency and interpretability of deepfake
detection models is essential for building
trust in their capabilities. Researchers are
1.Dataset Creation: Researchers have exploring methods to provide explanations
compiled diverse datasets containing both for model predictions and insights into the
real and fake videos to train deep learning features contributing to detection decisions.
models effectively. These datasets often
include samples from different sources and
cover a wide range of scenarios to enhance 5.Deployment and Scalability: Deploying
model generalization. deepfake detection systems at scale presents
challenges related to computational
resources, real-time processing, and
2.Feature Extraction: Feature extraction plays integration with existing platforms.
a crucial role in deepfake detection. Researchers are developing scalable solutions
Researchers have investigated methods for that can handle large volumes of data and
extracting meaningful features from videos, operate in real-time environments.
 cropped, resulting in a refined dataset
containing solely the facial components
Overall, the field of deepfake video detection
essential for analysis. To optimize
is interdisciplinary, drawing from computer
computational efficiency, a predetermined
vision, machine learning, signal processing,
threshold of 150 frames per video is
and cybersecurity. Ongoing research efforts
established, ensuring uniformity in frame
continue to advance the state-of-the-art in
count across the dataset.
deepfake detection, with the ultimate goal of
mitigating the societal risks associated with
the spread of manipulated media.
3. Dataset Split:

- Following data preprocessing, the dataset

is partitioned into distinct training and testing
subsets using a stratified 70-30 split. This
Implementation Details strategic division ensures that both the
training and testing sets maintain a balanced
The methodology employed in this project for representation of real and fake videos,
the detection of deepfake videos thereby facilitating robust model training and
encompasses a systematic approach involving evaluation.
various critical stages:

4. Model Architecture:
1.Dataset Compilation:
- The cornerstone of our deepfake detection
- The foundation of our project lies in the framework lies in its innovative architecture,
compilation of a comprehensive dataset which synergistically integrates Convolutional
sourced from diverse repositories, including Neural Networks (CNNs) and Recurrent
prominent collections such as DFDC, Neural Networks (RNNs). Leveraging the
FaceForensic++, and Celeb-DF. This power of transfer learning, a pretrained
amalgamation ensures the inclusivity and ResNext CNN model is employed to extract
representativeness of the dataset, covering a intricate features at the frame level. These
wide spectrum of real-world scenarios. To extracted features are then fed into a
mitigate bias during model training, an equal meticulously crafted Long Short-Term
number of real and fake videos are Memory (LSTM) network, enabling sequential
meticulously selected, thereby fostering a analysis and classification. The sequential
balanced learning environment. nature of LSTM facilitates temporal analysis of
the videos, crucial for discerning subtle cues
indicative of deepfake manipulation.
2. Data Preprocessing:

- A pivotal phase in our methodology is data

preprocessing, wherein videos undergo a 5. Hyperparameter Tuning:
series of transformations to isolate pertinent - An essential aspect of model optimization
facial regions. This process commences with involves fine-tuning hyperparameters to
the division of each video into individual enhance performance. This entails the
frames, followed by the application of utilization of the Adam optimizer, with a
sophisticated facial detection algorithms to carefully selected learning rate of 1e-5 and
accurately identify and extract facial features. weight decay of 1e-3. The choice of cross-
Subsequently, the identified facial regions are entropy loss as the objective function further
augments model training, aligning with the
classification nature of the deepfake detection
Results
task. We evaluated our algorithm on sequence
length of 10, 20,40,60,80,100.

By meticulously following this methodology,

our project endeavors to develop a robust
deepfake detection model capable of
effectively discerning between authentic and
manipulated videos, thereby safeguarding
against the proliferation of misinformation
and preserving the integrity of digital content.

The above image represents the results

achieved on our dataset by the model. The
accuracy in the image depicts the test
accuracy. As we can observe in our results that
the accuracy of the model is increasing with
the increasing number of sequence lengths.

Based on our results we can say that, our

model is able to predict whether the video is a

deepfake or real by seeing just 10 frames i.e.

less than 1 second (considering 30 frames per

second video) with a

decent accuracy of 84%.