© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.

org (ISSN-2349-5162)

DEEP-FAKE DETECTION:
MACHINE LEARNING APPROACHES TO COMBAT DEEP-FAKE
THREATS
1
Velpooru Venkata Sai Thushar, 2Rupani Varun Shiva Krishna, 3Kondadi Tejith
1,2,3
Students
Department of CSE, Institute of Aeronautical Engineering, Dundigal, Hyderabad-500043, Telangana, India

ABSTRACT

The term deep-fake emerged towards the end of 2017 within the reddit community a popular an American social media platform focused
on content rating and discussions since then deep-fake technology has rapidly advanced enabling the crafting incredibly lifelike media
material that convincingly mimics individuals appearances however its potential for misuse including the spread of disinformation and
deception of users has sparked major apprehensions in response researchers have been actively exploring novel approaches to develop
effective deep-fake detection systems this summary provides an overview of the latest advancements in employing machine learning
techniques for deep-fake detection aiming to mitigate the risks associated with its misuse.

The proposed approach trains machine learning models on a broad dataset that contains both real and deep-fake material it uses essential
characteristics taken from eye movements and facial expressions to distinguish between real and fake information the focus of the study
is on using deep neural networks and making sure that they can adapt to new developments in deep-fake generating technology.

To check the model's precision, accuracy, and recall validation and testing are done on separate datasets. Ethical questions linked to
privacy and consent are important to the study framework, addressing concerns connected with the analysis of media material. The
research stresses the continual improvement of detection algorithms, including updates to defeat developing deep fake tactics efficiently.

The outputs of this research seek to contribute to the development of practical and ethical deep fake detection technologies. As the
threat landscape advances, it is vital to be at the forefront of technical innovations to prevent the malicious use of synthetic media.
This research gives useful information for the continuing attempts to limit the hazards related with deep fake propagation.

Keywords: Deepfake Detection, Multi-task Cascaded Convolutional Networks (MTCNN), InceptionresnetV1, GradCAM, Gardio

JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a231
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

1) Understanding the Deep-fake Phenomenon: By examining

1. INTRODUCTION the evolution and proliferation of deep-fake technology, this
research seeks to gain insight into the underlying mechanisms
In recent years the advancement of deep learning technology driving its growth and impact on society.
has permitted the fabrication of hyper-realistic audiovisual
material known as deep fakes these sophisticated 2) Strengthening Detection Systems: Through the exploration
manipulations powered by artificial intelligence algorithms of machine learning approaches, this study aims to enhance
have sparked considerable concerns owing to their potential the effectiveness of deep-fake detection systems by
for destructive purposes like the dissemination of leveraging diverse datasets and key elements extracted from
misinformation and the weakening of confidence in digital facial expressions, eye movements, and speech patterns.
media as the borders between truth and fabrication blur the
3) Ethical Considerations: Ethical questions related to
requirement to build effective deep fake detection algorithms
privacy, consent, and the responsible use of deep-fake
becomes increasingly clear this research article investigates
detection technologies are integral to this research. By
the environment of deep fake detection using machine
addressing these concerns, we strive to ensure that our
learning seeking to contribute to the creation of effective and
methodologies adhere to ethical principles and safeguard
adaptable solutions in the face of developing synthetic media
individual rights.
threats deep fakes driven by generative adversarial networks
GANS and deep neural networks have been increasingly 4) Continual Improvement: Recognizing the dynamic nature
available permitting anybody of the deep-fake landscape, this research emphasizes the
importance of continual improvement in detection
with modest technological competence to make convincing
algorithms. By staying ahead of evolving deep-fake tactics,
forgeries of audio video and even literary content with
we aim to develop detection systems that are resilient to
repercussions for several fields like politics entertainment
emerging threats.
and cybersecurity the necessity to prevent the harmful use of
deep fakes has gotten crucial the beginning of this study
article describes the importance of the deep fake phenomena
stressing its prospective societal effect secondly it
underscores the requirement to strengthen detection systems
to ensure the integrity of digital information as we traverse a
future dominated by ai-generated content recognizing and 2. DEEP-FAKE GENERATION
minimizing the dangers associated with deep fakes is vital to
sustaining confidence in the digital domain this project seeks 3.1 How deep-fakes are produced
to add to this critical conversation by exploring and proposing Deep-fake models are typically created using deep learning
unique ways for the reliable identification of deep fakes via techniques, particularly generative adversarial networks
the perspective of machine learning.BACKGROUND (GANs) and autoencoder-based architectures.

1) Data Collection: Deep-fake models require large datasets

2.1 An overview of deep-fake of images or videos of the target person(s) to be manipulated.
These datasets serve as the basis for learning the facial
Deep-fake technology, a term coined in late 2017 by a Reddit features, expressions, and mannerisms that the model will
user, marks a pivotal moment in the evolution of digital replicate.
manipulation. Initially surfacing as a niche phenomenon,
deep-fakes swiftly infiltrated mainstream discourse, 2) Preprocessing: The collect data is preprocessed to extract
propelled by advancements in deep learning algorithms and facial landmarks, align faces, and normalize lighting
the widespread availability of user-friendly software. This conditions. Preprocessing helps ensure consistency and
surge in accessibility democratized the creation of hyper- quality in the training data.
realistic manipulated media, allowing individuals with 3) Training the Generative Model: Generative adversarial
minimal technical proficiency to fabricate convincing networks (GANs) are commonly used for deep-fake
audiovisual content. Consequently, concerns regarding the generation. GANs consist of two neural networks: a
misuse of deep-fakes have escalated, fueled by instances of generator and a discriminator. The generator learns to
their exploitation for nefarious purposes such as spreading generate realistic fake images/videos from random noise,
false information, defaming public figures, and manipulating while the discriminator learns to distinguish between real and
public opinion. The proliferation of deep-fakes poses a fake images/videos.
profound threat to the fabric of trust underpinning societal
discourse and democratic processes, casting doubt on the 4) Training Process: The generator generates fake
authenticity of digital content and eroding confidence in images/videos using random noise as input. The
established institutions. As the boundaries between reality discriminator is trained to differentiate between real and fake
and fabrication blur, the imperative to develop robust images/videos. The generator and discriminator are trained
detection mechanisms to combat the insidious spread of iteratively in an adversarial manner. The generator tries to
deep-fakes becomes increasingly urgent. This backdrop generate increasingly realistic images/videos to fool the
underscores the critical importance of research endeavors discriminator, while the discriminator tries to become better
aimed at devising innovative and effective solutions to detect at distinguishing between real and fake images/videos.
and mitigate the impact of deep-fake manipulation on society.
5) Loss Function: The training process is guided by a loss
This study aims to address several key objectives: function that measures how well the generator is fooling the
JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a232
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

discriminator. Common loss functions include binary cross- and lighting reflections, especially in glasses, ensuring they
entropy or Wasserstein distance. align with movement. Examine facial hair, moles, and
blinking patterns for authenticity. Pay close attention to lip
6) Fine-Tuning and Refinement: After the initial training, the movements, as discrepancies in syncing may indicate
model may undergo fine-tuning and refinement to improve manipulation. Overall, meticulous observation of these cues
the quality of generated deep-fakes. This may involve can help detect subtle anomalies indicative of deep-fake
adjusting hyperparameters, using larger datasets, or alterations.
employing advanced techniques such as self-attention
mechanisms. These are few parameters that we can look for to detect deep-
fakes.
7) Post-Processing: The generated deep-fakes may undergo
post-processing to enhance their realism. This could include
adding imperfections, adjusting lighting, or smoothing
transitions between frames.

8) Deployment: Once trained, the deep-fake model can be 4.2 Deep-fake detection mathematical model?
deployed to generate manipulated images or videos. These
deep-fakes can be shared online or used for various purposes, The mathematical model we used involves several
ranging from entertainment to malicious deception. components:

a) Face Detection with MTCNN: Multi-task Cascaded

Convolutional Networks (MTCNN) is used for face
detection. MTCNN is a framework developed as a solution
for both face detection and face alignment. The MTCNN
model detects faces in an image by applying a series of
convolutional filters to identify facial landmarks and
bounding boxes.

Fig2: Process of MTCNN

b) Face Recognition with InceptionResnetV1: The

InceptionResnetV1 model is used for face recognition. This
model takes a cropped face image as input and outputs a
Fig1: Deep-fake Generation Model feature vector that represents the characteristics of the face.

b.1) How face recognition with InceptionResnetV1

typically works?
3. DEEP-FAKE DETECTION
 Model Structure: Inceptionresnetv1 is built upon a
4.1 How can we detect deep-fakes?
framework of convolutional-pooling and fully-
According to MIT media lab there are several Deep-fake connected-layers it integrates inception-modules
artifacts that we can be on the lookout for. When scrutinizing enabling parallel processing of different receptive
for deep-fakes, focus on facial features, particularly the eyes, field sizes within each network layer ultimately
cheeks, and forehead, assessing for inconsistencies such as boosting its performance capabilities.
overly smooth or wrinkled skin. Check for natural shadows
JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a233
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

 Pre-trained parameters: The pretrained- using the embedding for tasks such as distance calculation or
parameters of InceptionResnetV1 often involve classification.
weights learned from vast datasets containing facial
c) GradCAM for Explainability: GradCAM (Gradient-
images. This pretraining phase equips the model
weighted Class Activation Mapping) is used to visualize the
with the capability to extract distinct facial features,
regions of the input image that are most influential in the
which proves beneficial for tasks related to
model's decision-making process. It computes the gradients
individual recognition.
of the target class output with respect to the feature maps of
 Feature-Extraction: A face-containing input photo
the specified layers in the model. This helps in understanding
is sent through the network in order to identify a face
which parts of the image contribute the most to the model's
using InceptionResnetV1. Features at various
prediction.
degrees of abstraction are taken from the photo as it
moves through the network's layers. These features d) Sigmoid Activation for Classification: The output of the
record details including high-level traits, textures, model is passed through a sigmoid activation function, which
and face shapes. squashes the output values between 0 and 1. This output
 Embedding Generation: The outcome derived represents the confidence score for the input image being
from the ultimate layer of the network, prior to the classified as a real or fake face. Sigmoid activation function
classification layer, is frequently utilized as a feature is used in classification:
vector or embedding to portray the input face. This
embedding encapsulates essential details about the  Output Interpretation: The output of the sigmoid
facial traits crucial for recognition purposes. function represents the probability or confidence
 Distance Metric: When comparing two face score that an input belongs to the positive class (e.g.,
embeddings to ascertain their potential belonging to class 1). For binary classification, this is often
the same individual, a distance metric such as interpreted as the probability of the input belonging
Euclidean distance or cosine similarity is frequently to the "positive" class.
employed. Smaller distances imply higher similarity  Decision Boundary: In binary classification, a
between embeddings, indicating a higher likelihood decision boundary is typically set at 0.5. Inputs with
that the faces originate from the same person. sigmoid output greater than 0.5 are classified as
 Classification (Optional): In some instances, the belonging to the positive class, while inputs with
output from the embedding-layer might undergo output less than 0.5 are classified as belonging to the
processing via a classification-layer to execute tasks negative class.
like face verification ascertaining if two faces  Training: During the training process, the sigmoid
pertain to the same individual or face identification function is applied to the output of the last layer of
assigning identities to faces. the neural network. The parameters of the network
(weights and biases) are adjusted using optimization
algorithms such as gradient descent to minimize the
difference between the predicted outputs and the
b.2) Mathematical model for face recognition with
true labels of the training data.
InceptionResnetV1:
 Loss Function: The output of the sigmoid function
1. Input Image: We have an input image (I). is often used as input to a loss function, such as
binary cross-entropy loss, which quantifies the
2. CNN Operations: The image is processed through a deep difference between the predicted probabilities and
neural network (InceptionResnetV1) to extract features: the true labels. This loss is minimized during
F = CNN(I) training to improve the model's ability to correctly
classify inputs.
where F represents the features extracted from the image.

3. Embedding Generation: The features F are converted into

a lower-dimensional representation or embedding (E):

E = Embed(F)

4. Distance Metric: If needed, the distance between two

embeddings E1 and E2 can be computed:

d(E1,E2)=|E2-E1|

5. Classification: Optionally, the embedding E can be used Fig3: Sigmoid function graph
for classification:
e) Confidence Scores: The confidence scores for real and
P = Classify(E) fake faces are computed based on the sigmoid output. The
score closer to 1 indicates higher confidence in the predicted
where P represents the probability distribution over classes. class. The confidence scores for classifying whether the input
In summary, the simplified mathematical model involves image contains a real or fake face are calculated using a
processing the input image through a CNN to extract features, sigmoid activation function applied to the output of the face
converting these features into an embedding, and optionally recognition model. Here's how we have computed:

JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a234
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

4. Deep-fake working model

5.1) Real image Input:
The input to the deep-fake detection model is an image
containing a face. This image is expected to be in a format
compatible with the gradio library's Image input type, which
typically accepts PIL (Python Imaging Library) Image
objects. Here are the details of the input:
 Type: Image (PIL format)
 Content: The input image should contain a single
Fig4: Code for confidence Scores face that needs to be classified as real or fake.

Output contains the sigmoid-activated output of the face

recognition model. The sigmoid function squashes the raw
output values to the range= [0, 1], representing the
confidence scores for each class (real or fake).

 Real_prediction: represents the confidence score for

the image being classified as "real" face.
 Fake_prediction: represents the confidence score for
the image being classified as "fake" face.

f) Image Processing: Various image processing techniques

such as resizing, normalization, and overlaying the
GradCAM heatmap on the original image are used for
visualization and interpretability. Image processing steps
Fig5: Example input image.
involved:

5.2) Output for real images:

Face Detection Preprocessing:
The output of the deep-fake detection model consists of two
 The input image is passed through the MTCNN components: a label indicating whether the face is predicted
(Multi-task Cascaded Convolutional Networks) as real or fake, and an image showing the input face with an
model for face detection. overlaid heatmap for explanation. Here are the details of the
 The detected face region is cropped and resized to a output:
fixed size (256x256) using bilinear interpolation. Label:
Visualization:  Type: String
 Content: The predicted label indicating whether the
 The preprocessed face image is converted to a
input face is classified as "real" or "fake".
numpy array for visualization.
 Before passing the face image to the model for Face with Explainability:
prediction, it's normalized by dividing by 255.0 to
 Type: Image (PIL format)
scale pixel values between 0 and 1.
 Content: An image showing the input face with an
GradCAM Visualization: overlaid heatmap generated by GradCAM. This
heatmap highlights the regions of the face that are
 GradCAM is applied to the preprocessed face image most influential in the model's decision-making
to generate a heatmap highlighting the important process, providing an explanation for the prediction.
regions for the model's prediction.
 The heatmap is overlaid onto the original face image
to visualize which parts of the face are contributing
the most to the model's decision.
 This overlay is created using the cv2.addWeighted
function, blending the original face image with the
heatmap.
Fig6: Predicted label indication for fig5.

It says that the input image is 100% real and 0% fake.

Therefore, we can confidently say that the input image is a
legit image not a deep-fake image.

JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a235
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

5. Applications of deep-fake detections

 Media Integrity: Helps in detecting false or deep-

faked content. It helps in ensuring trust in the social
media and prevents the spread of false content.
Suppressing deep-fakes results in achieving the
paramount, authentic information.

 Cybersecurity: Deep-fake detection can be utilized

in cybersecurity to prevent malicious actors from
creating fake identities or impersonating individuals
through manipulated media. This can help in
mitigating the risks associated with identity theft,
Fig7: Input face with an overlaid heatmap generated by phishing attacks, and social engineering scams.
GradCAM.
These were the predictions made my our deep-fake detection  Forensic Analysis: Deep-fake detection techniques
model. The photo is taken from one of his social media can assist forensic analysts and law enforcement
accounts. agencies in verifying the authenticity of digital
evidence such as images, videos, and audio
5.3) Deep-fake Image Input: recordings. This is essential in investigations and
legal proceedings to ensure that evidence presented
is genuine and not tampered with.
 Political Disinformation and Election Integrity:
Deep-fake detection technologies play a crucial role
in combating political disinformation and
safeguarding election integrity. By identifying and
flagging manipulated media content, these tools can
help prevent the spread of misinformation and
manipulation aimed at influencing public opinion or
election outcomes.

 Content Moderation: Social media platforms and

online communities can use deep-fake detection to
Fig8: An AI generated image moderate content and prevent the spread of harmful
or deceptive media. This helps maintain a safe and
5.4) Output images for fig8: trustworthy environment for users and reduces the
dissemination of misleading or malicious content.

 Privacy Protection: Deep-fake detection tools can

also be used for protecting individuals' privacy by
identifying and removing unauthorized or
maliciously manipulated images shared without
one’s knowledge.

Fig9: Predicted label indication.

It says that the input image is 100% fake and 0% real. 6. Conclusion:
Therefore, we can confidently say that the input image is a In conclusion, this documentation provides a comprehensive
not legit image not a deep-fake image. overview of the deep-fake phenomenon, its implications, and
the efforts to develop effective detection techniques using
machine learning approaches. Deep-fake technology, driven
by advances in deep learning algorithms, has raised
significant concerns due to its potential for spreading
misinformation and undermining trust in digital media.
The documentation outlines the process of deep-fake
generation, highlighting the use of generative adversarial
networks (GANs) and deep neural networks to create hyper-
realistic manipulated media. It also discusses key indicators
and artifacts that can be used to detect deep-fakes,
emphasizing the importance of paying attention to facial
features, expressions, and inconsistencies in the generated
content.
Fig10: Input face with an overlaid heatmap generated by
GradCAM. Furthermore, the documentation delves into the mathematical
models and techniques employed for deep-fake detection,
including face detection with MTCNN, face recognition with
InceptionResnetV1, GradCAM for explainability, and
JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a236
© 2024 JETIR March 2024, Volume 11, Issue 3 www.jetir.org (ISSN-2349-5162)

sigmoid activation for classification. These models, coupled

with image processing techniques, enable the visualization
and interpretation of deep-fake detection results.
The applications of deep-fake detection span various
domains, including media integrity, cybersecurity, forensic
analysis, political disinformation, brand protection, and
content moderation. By leveraging deep-fake detection
technologies, stakeholders can mitigate the risks associated
with the malicious use of synthetic media and safeguard the
authenticity of digital content.
Overall, this documentation underscores the importance of
ongoing research and innovation in deep-fake detection to
address the evolving threat landscape and preserve trust in
digital communication channels. By advancing detection
capabilities and addressing ethical considerations, we can
better protect individuals, organizations, and society as a
whole from the harmful effects of deep-fake manipulation.

JETIR2403030 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org a237