Page 792 - Emerging Trends and Innovations in Web-Based Applications and Technologies

International Journal of Trend in Scientific Research and Development (IJTSRD)
Special Issue on Emerging Trends and Innovations in Web-Based Applications and Technologies
Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470

Deepfake Detection with Generalization
via Domain-Aware Meta-Learning

2
Kartik. S. Raut. , Harsh. S. Pendke , Prof. Anupam Chaube , Prof. Usha Kosarkar
4
1
3
1,2,3,4 Department of Science and Technology,
1,2 G H Raisoni Institute of Engineering and Technology, Nagpur, Maharashtra, India
3,4 G H Raisoni College of Engineering and Management, Nagpur, Maharashtra, India

ABSTRACT performance against the latest deepfake techniques, which
Deepfakes, media manipulated using deep learning can evade current detection systems (Le et al., 2024).
techniques, pose a growing threat to the integrity of digital One promising approach to address the limitations in
content. These AI-generated forgeries are becoming
increasingly sophisticated, making them difficult to detect. generalizing deepfake detection involves utilizing a meta-
Traditional detection methods often lag behind the rapid learning framework for domain generalization, combined
evolution of deepfake techniques and are hampered by the with data augmentation. Meta-learning enhances
limited variety of training data, making it hard for them to adaptability by exposing the model to a variety of
manipulations during training, enabling it to learn features
generalize effectively to new types of deepfakes. This thesis
introduces a novel deepfake detection approach that that remain consistent across different deepfake generation
combines meta-learning for domain generalization (MLDG) techniques. Meanwhile, data augmentation increases the
diversity of training data, further strengthening the model’s
with self blended images (SBI) to address this challenge.
MLDG, inspired by meta-learning principles, aims to ability to detect unseen deepfakes.
improve the model’s adaptability to new manipulation This study evaluates the effectiveness of an integrated
techniques by simulating domain shifts during training. The deepfake detection framework that combines Meta-
model learns from various source domains representing Learning for Domain Generalization (MLDG) (D. Li et al.,
different deepfake generation methods. Additionally, SBIs, 2017b) with Self-Blended Images (SBIs) (Shiohara &
synthetic images created by blending real and manipulated Yamasaki, 2022) as a data augmentation strategy. The
faces, are incorporated to further diversify the training data primary objective is to assess how well this approach
and promote the learning of features that generalize across enhances detection performance on previously unseen
domains. This thesis focuses on detecting image-based deepfake generation techniques compared to a conventional
deepfakes using the Face Forensics++ dataset, a benchmark baseline.
collection of real and manipulated videos, specifically
The evaluation will be conducted using the FaceForensics++
designed for deepfake detection research. The proposed
dataset (Rössler et al., 2019) through a leave-one-out
method is evaluated with a leave-one-out cross-validation
cross-validation methodology. The model will be trained on
scheme on this dataset, where each deepfake generation
multiple deepfake generation techniques and tested on
technique is used as a test case while the others are used for
unseen methods. The performance of the proposed approach
training. The results consistently show that MLDG, when
will be compared against a baseline model trained using
enhanced with SBIs, outperforms the standard Empirical Empirical Risk Minimization (ERM) (Vapnik, 1999) to
Risk Minimization (ERM) method, demonstrating its
determine its effectiveness.
effectiveness in generalizing to unseen manipulation
techniques. The research offers a practical solution for This study focuses exclusively on video-based deepfake
deepfake detection, highlighting how MLDG and SBI detection, specifically targeting facial manipulations. While
augmentation can create more effective and adaptable audio-based deepfake detection is an emerging concern, it is
detection systems. The findings emphasize the need for beyond the scope of this research. Additionally, the detection
models that can adapt to evolving deepfake techniques to strategy used in this study is image-based, analyzing
protect the integrity of digital media. individual video frames instead of incorporating temporal

information. This allows for a focused evaluation of domain
I. INTRODUCTION generalization techniques applied to facial deepfake
The rapid advancements in deepfake generation technology detection.
have introduced significant challenges for detection II. RELATED WORK
methods, highlighting the need for more robust and
Challenges in Traditional Machine Learning for
adaptable approaches. Traditional deepfake detection relies
Deepfake Detection
on machine learning models trained to distinguish real and
Conventional machine learning models face several
manipulated media based on specific deepfake techniques.
limitations, particularly their dependence on large labeled
However, these models often struggle when encountering
datasets for effective learning. In real-world deepfake
previously unseen manipulation methods (Malik et al.,
detection, obtaining labeled datasets for every possible
2022). Recent research efforts have aimed to improve the
manipulation technique is challenging, time-consuming, and
adaptability of detection models across different types of
often impractical (Prince, 2023). Additionally, machine
deepfake manipulations (Sun et al., 2021; Z. Wang et al.,
learning models frequently struggle with adapting to
2023), yet there remains a notable gap in model
evolving or unseen tasks. For example, facial recognition
IJTSRD | Special Issue on Emerging Trends and Innovations in Web-Based Applications and Technologies Page 782