Studi komparatif model cnn untuk pengenalan facial anomalies pada citra deepfake

Siahaan, Matthew (2025) Studi komparatif model cnn untuk pengenalan facial anomalies pada citra deepfake. Bachelor thesis, Universitas Pelita Harapan.

[thumbnail of Title] Text (Title)
Title.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (114kB)
[thumbnail of Abstract] Text (Abstract)
Abstract.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (412kB)
[thumbnail of Chapter1] Text (Chapter1)
Chapter1.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (466kB)
[thumbnail of Chapter2] Text (Chapter2)
Chapter2.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (1MB)
[thumbnail of Chapter3] Text (Chapter3)
Chapter3.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (954kB)
[thumbnail of Chapter4] Text (Chapter4)
Chapter4.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (2MB)
[thumbnail of Chapter5] Text (Chapter5)
Chapter5.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (281kB)
[thumbnail of Bibliography] Text (Bibliography)
Bibliography.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (313kB)
[thumbnail of Appendices] Text (Appendices)
Appendices.pdf
Restricted to Repository staff only
Available under License Creative Commons Attribution Non-commercial Share Alike.

Download (1MB)

Abstract

Teknologi deepfake yang semakin berkembang telah menimbulkan tantangan besar dalam aspek keaslian informasi dan kepercayaan publik terhadap konten digital. Deepfake memungkinkan manipulasi citra wajah manusia secara realistis dengan menggunakan model berbasis kecerdasan buatan (Artificial Intelligence, AI). Oleh karena itu, diperlukan sistem deteksi otomatis yang mampu mengenali facial anomalies dalam citra deepfake secara akurat dan efisien. Penelitian ini bertujuan untuk membandingkan performa antara dua arsitektur Convolutional Neural Network (CNN), yaitu MesoNet (Meso-4 dan MesoInception-4) dan ResNet (ResNet18 dan ResNet50) dalam mendeteksi facial anomalies pada citra deepfake. Dua jenis dataset digunakan dalam eksperimen, yaitu dataset publik “140K Real and Fake Faces” dan dataset mandiri yang terdapat tambahan citra hasil manipulasi mandiri menggunakan Grok AI. Model diuji berdasarkan metrik akurasi, presisi, recall, F1-score, dan waktu pemrosesan. Hasil penelitian menunjukkan bahwa ResNet18 memberikan performa terbaik dengan F1-score yang paling tinggi dan waktu pemrosesan yang cukup efisien. MesoInception-4 menjadi alternatif ringan dengan akurasi yang cukup baik. Penelitian ini menunjukkan bahwa pemilihan model harus mempertimbangkan trade-off antara akurasi dan efisiensi komputasi, serta pentingnya dataset yang bervariasi dalam meningkatkan performa model. / The advancement of deepfake techonology has raised significant challenges in the authenticness of information and public trust toward digital content. Deepfake enables realistic manipulation of human facial images using artificial intelligence-based models. Therefore, it is crucial to develop an automatic detection system that can accurately and efficienly recognize facial anomalies in deepfake images. This study aims to compare the peformance of two Convolutional Neural Network (CNN) architectures, namely MesoNet (Meso-4 and MesoInception-4) and ResNet (ResNet18 and ResNet50), in detecting facial anomalies in deepfake images. Two types of datasets are used in the experiment: the public dataset “140K Real and Fake Faces” and a curated dataset that includes additional self-manipulated deepfake images generated using Grok AI. The models were tested based on accuracy, precision, recall, F1-score, and processing time metrics. The results showed that ResNet18 gave the best performance with the highest F1-score and efficient processing time. MesoInception-4 is a lightweight alternative with good accuracy. This study shows that model selection should consider the trade-off between accuracy and computational efficiency, and the importance of varied datasets in improving model performance.
Item Type: Thesis (Bachelor)
Creators:
Creators
NIM
Email
ORCID
Siahaan, Matthew
NIM03082210043
matthewfabio99@gmail.com
UNSPECIFIED
Contributors:
Contribution
Contributors
NIDN/NIDK
Email
Thesis advisor
Yugopuspito, Pujianto
NIDN0324086701
UNSPECIFIED
Uncontrolled Keywords: Deepfake; CNN; ResNet; MesoNet; Facial Anomalies; Deteksi Wajah
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Depositing User: Matthew Fabio Orlando S
Date Deposited: 22 Jul 2025 04:12
Last Modified: 22 Jul 2025 04:12
URI: http://repository.uph.edu/id/eprint/69834

Actions (login required)

View Item
View Item