Department of Computer Science and Engineering

“Development of AI/ML based solution for detection of face-swap based

deep fake videos ”
Under the Guidance of “Prof. Rahul kulkarni”
Team members :
Abhishek.ais (3GN23CS400) Akash kumar (3GN23CS402)
Chidanand (3GN23CS403) Manikanth (3GN23CS405)
INTRODUCTION
•Deepfake technology uses AI to manipulate videos, with face-swapping being one of the
most common and dangerous techniques.

•These videos are used to spread misinformation, harass individuals, and impersonate
identities.

•The realistic nature of face-swapped videos makes them hard to detect with the human eye.

•Therefore, developing an AI/ML-based detection system is essential to identify and curb their
misuse.
Motivation

•Rise in the number of deepfake incidents affecting individuals, celebrities, and

political figures.

•Lack of reliable tools available to the public and organizations for verifying video
authenticity.

•Potential for misuse in legal, financial, and social domains if left unchecked.

•Encouragement to contribute to digital media forensics through a technical solution.

 Literature Review
KeyKeyContributions
S.No.
S.No. Title
Title Authors
Authors Year
Year Method
Method/ /Model
Model Dataset(s)
Dataset(s) / Findings
Contributions /
Findings
Established a strong
FaceForensics++: CNNEstablished
baseline for a
Learning to Detect strong CNN
detecting facial
1 FaceForensics++: Rossler et al. 2019 XceptionNet (CNN) FaceForensics++
Manipulated Facial baseline for
manipulations,
Learning to Detect
1 Images Rossler et al. 2019 XceptionNet (CNN) FaceForensics++ especially
detecting facial
face-
Manipulated Facial
manipulations,
swaps.
Images
especially face-
Designed
swaps. a
MesoNet: A
lightweight network
Compact Facial MesoNet (Shallow Self-generated
2 Afchar et al. 2018 Designed
effective for alow-
MesoNet: A
Video Forgery CNN) dataset
lightweight
resolution video
Compact
DetectionFacial
Network MesoNet (Shallow Self-generated
2 Afchar et al. 2018 network effective
deepfakes.
Video Forgery CNN) dataset
for low-resolution
Introduced a
Detection Network
FakeCatcher: video deepfakes.
method using blood
Detection of
PPG-based analysis flow signals toa
Introduced
3 Synthetic Portrait
FakeCatcher: Ciftci et al. 2020 In-house dataset
(physiological) detect
methoddeepfakes,
using
Videos using
Detection of PPG-based robust to
blood flow signals
3 BiologicalPortrait
Synthetic Signals Ciftci et al. 2020 analysis In-house dataset compression.
to detect
Videos using (physiological) Used frequency
DeepFakeSignals
Detection deepfakes, robust
Biological analysis and deep
by Analyzing FaceForensics++, to compression.
4 Frank et al. 2020 Frequency + CNN learning to uncover
Convolutional Celeb-DF Used frequency
DeepFake generator-specific
Traces analysis and deep
Detection by fingerprints.
FaceForensics++, learning to
4 Analyzing Frank et al. 2020 Frequency + CNN Summarized
Celeb-DF uncover generator-
Convolutional existing deepfake
specific
A Survey on
Traces detection methods
Survey (CNN, RNN, fingerprints.
5 Deepfake Detection Tolosana et al. 2020 Multiple datasets and categorized
GAN)
Techniques them by input
Summarized
modality
existinganddeepfake
A Survey on
Problem Identification
•Most existing systems fail to detect deepfake videos that are edited with high precision and
low compression.

•There is a lack of lightweight and real-time detection tools that can be deployed on mobile or
edge devices.

•Deepfake detection becomes difficult when the video quality is low or contains background
noise.

•Current models often perform well only on specific datasets and fail to generalize to real-
world videos
•.
•Many detection tools are not user-friendly or accessible to non-technical users, limiting their
widespread use.
Objectives

• To design and implement an AI/ML-based system for detecting face-swap deepfake videos.

•To use deep learning models trained on labeled datasets for high accuracy.

•To evaluate the model’s performance using precision, recall, F1-score, and accuracy metrics.

•To deploy a user-friendly interface for uploading and analyzing video content.
DESIGN
Flowchart
•Hardware Requirements:
1. Local Development Machine (Minimum)

Component Recommended
CPU Intel i7 / AMD Ryzen 7 or better
RAM 32 GB (16 GB minimum)
NVIDIA RTX 3060+ (minimum
GPU
8GB VRAM)
SSD 1TB (especially for high-
Storage
speed video I/O)
2. For Model Training (if local)

Component Recommended
NVIDIA RTX 3090 / A6000 or
GPU
better (24–48 GB VRAM)
RAM 64–128 GB
For training large models or
Multi-GPU Support
multi-stream video processing
Software Requirements:
1. Operating System
•Preferred: Ubuntu 20.04+ (Linux is better for ML workflows)
•Alternative: Windows 10/11 (with WSL2 if needed)

2. Programming Languages
•Python 3.8+ (most libraries and frameworks are Python-based)

3. Libraries and Frameworks

ML/DL Frameworks:
•TensorFlow 2.x or PyTorch 2.x (choose one based on your expertise)
•scikit-learn – for classical ML tasks and metrics
•OpenCV – for video and image processing
•dlib or face_recognition – for face detection and alignment
•MMDetection or YOLOv8 (optional) – for robust face/object detection
4. Development Tools
•Jupyter PyCharmNotebook / JupyterLab
•VS Code /
•Docker – for reproducibility
•Git – version control
•MLflow / Weights & Biases / TensorBoard – for experiment
tracking

5. Databases & Storage

•PostgreSQL or MongoDB – for metadata storage
•Amazon S3 / Google Cloud Storage – for large-scale video
storage
Conclusion
•Deepfake detection is a critical area in the AI security landscape.

•This project contributes by offering a practical AI/ML solution to detect face-swapped

deepfakes with good accuracy.

•With further development, the model can be integrated into content verification platforms and
media monitoring tools.
References:

1.Rössler, A. et al. (2019). FaceForensics++: Learning to Detect Manipulated

Facial Images. arXiv preprint arXiv:1901.08971.

2.Li, Y. et al. (2020). Celeb-DF: A Large-scale Challenging Dataset for

DeepFake Forensics. IEEE CVPR.

3.Dolhansky, B. et al. (2020). The DeepFake Detection Challenge Dataset.

arXiv:2006.07397.

4.Afchar, D. et al. (2018). MesoNet: a Compact Facial Video Forgery

Detection Network. WIFS 2018.
THANK YOU