Department Of CSE-AI/ML& IOT
ADITYA COLLEGE OF ENGINEERING AND TECHNOLOGY(A)
DEEP FAKE
DETECTION
The Eye of Authenticity
Internship Coordinator: Presented by:
Mrs. [Link] Jabili Devi Ratnala Raja
[Link], ([Link]) 21MH1A4949
Assistant Professor
INTRODUCTION
• DeepFake Detection is the task of detecting fake videos
or images that have been generated using deep learning
techniques.
• DeepFakes are created by using Deep learning
algorithms to manipulate or replace parts of an original
video or image, such as the face of a person.
• The goal of deepfake detection is to identify such
manipulations and distinguish them from real videos or
images.
Reasons for pursuing this project !
• We chose this project to address the challenge of detecting the truth behind IMAGES/VIDEOS.
• We couldn't find any websites that detect whether an IMAGE/VIDEO is REAL or FAKE, so we
decided to create one ourselves.
• Through our website, we will detect whether the uploaded data is real or fake. This will enable
people to discern the truth about the data they encounter.
Data Insights
• The dataset, obtained from the Kaggle website, occupies 10.16 GB
of storage space.
• It comprises three folders categorized as:
1. Celeb – YouTube – real : 712 videos
2. Celeb – YouTube – fake : 5305 videos
3. Test : 518 videos
• For convenience, we have downsized the dataset to 358 MB in size.
• This reduced dataset now consists of 250 videos.
1. Celeb – YouTube – real : 100 videos
2. Celeb – YouTube – fake : 100 videos
3. Test : 50 videos
Architecture
• We have developed a system to demonstrate the analysis and differentiation of authentic and
manipulated media.
• The process involves the following steps:
1. Data Collection: Initially, we gather a comprehensive dataset containing both authentic and
synthetic (deep fake) media.
2. Preprocessing: These includes segmentation, resizing, and augmentation, are applied to
optimize the data for training purposes.
3. Feature Extraction: This process utilizes techniques like CNN to identify distinctive patterns and
traits specific to manipulated content.
4. Model Training: Once the features are extracted, a Deep learning model is trained to classify input
media as either real or fake by using InceptionResNetV2(pre-trained) model.
5. Evaluation and Validation: Assessment is based on the probability of frames extracted from a video.
Our model undergoes validation using the "Test" folder.
6. User Interface: For simplicity, we have developed a website as the interface for users to upload their
data.
7. Deployment: The website has been made public by hosting it on the AWS cloud. Users can access it
using our IP address.
Library Suites:
Pre-Trained Model: Libraries:
• InceptionResNetV2 • NumPy
• Pandas
• Sklearn
• Tensorflow
• Keras
• OpenCV
• DLIB
Advantages:
• Maintaining trust in media and society
• Ensuring the safety of women
• Optimizing the time to uncover the truth
• Safeguarding personal privacy
• Combatting Misinformation
Conclusion:
• Deepfake detection is a critical area of research and development that plays a crucial role
• Helps to combat against the spread of misinformation, protecting individuals' privacy, and
preserving trust in media and society.
Thank You
