PENGEMBANGAN SISTEM DETEKSI VIDEO DEEPFAKE BERBASIS WEB MENGGUNAKAN PERBANDINGAN ALGORITMA CONVOLUTIONAL NEURAL NETWORK DAN TRANSFORMERS BERDASARKAN KETIDAKSEJAJARAN SUBPIKSEL

Muhammad Fariduddin Athar, - and Mahmudah Salwa Gianti, - and Dewi Indriati Hadi Putri, - (2025) PENGEMBANGAN SISTEM DETEKSI VIDEO DEEPFAKE BERBASIS WEB MENGGUNAKAN PERBANDINGAN ALGORITMA CONVOLUTIONAL NEURAL NETWORK DAN TRANSFORMERS BERDASARKAN KETIDAKSEJAJARAN SUBPIKSEL. S1 thesis, Universitas Pendidikan Indonesia.

Abstract

Penelitian ini bertujuan untuk mengembangkan sistem deteksi video deepfake berbasis web dengan memanfaatkan analisis ketidaksejaran subpiksel melalui optical flow serta membandingkan kinerja Convolutional Neural Networks (CNN) dan Transformers. Metode penelitian dilakukan secara ADDIE dengan tahapan preprocessing video melalui ekstraksi frame dan pemetaan fitur subpiksel, kemudian dilanjutkan dengan klasifikasi menggunakan arsitektur CNN dan Transformers. Evaluasi kinerja dilakukan berdasarkan akurasi, loss, recall, dan F1-score untuk menentukan model terbaik. Hasil penelitian menunjukkan bahwa CNN unggul dalam akurasi akurasi 98%, sementara Transformers hanya memiliki akurasi sekitar 50% saja. Analisis ini membuktikan bahwa integrasi optical flow mampu memperjelas perbedaan pola mikro antara video asli dan deepfake. Model dengan performa terbaik kemudian diimplementasikan ke dalam aplikasi web berbasis Streamlit, yang mendukung unggah video, analisis frame, dan deteksi real-time. Kesimpulannya, penelitian ini berhasil menunjukkan bahwa CNN masih memberikan akurasi lebih tinggi dibandingkan dengan Transformers dengan selisih yang cukup tinggi sehingga membuatnya lebih handal dalam mendeteksi video deepfake. ----- This study aims to develop a web-based deepfake video detection system by utilizing subpixel inconsistency analysis through optical flow and comparing the performance of Convolutional Neural Networks (CNN) and Transformers. The research method was conducted experimentally with video preprocessing stages through frame extraction and subpixel feature mapping, followed by classification using CNN and Transformer architectures. Performance evaluation was conducted based on accuracy, loss, recall, and F1-score to determine the best model. The results showed that CNN outperformed Transformers with an accuracy of 98.5%, while Transformers only achieved approximately 50% accuracy. This analysis demonstrates that integrating optical flow can clarify the differences in micro-patterns between original videos and deepfakes. The best-performing model was then implemented into a Streamlit-based web application that supports video uploads, frame analysis, and real-time detection. In conclusion, this study successfully demonstrated that CNN still provides higher accuracy compared to Transformers with a significant difference, making it more reliable in detecting deepfake videos.

	Text S_MKB_2103545_Title.pdf Download (706kB)
	Text S_MKB_2103545_Chapter1.pdf Download (304kB)
	Text S_MKB_2103545_Chapter2.pdf Restricted to Staf Perpustakaan Download (722kB)
	Text S_MKB_2103545_Chapter3.pdf Download (524kB)
	Text S_MKB_2103545_Chapter4.pdf Restricted to Staf Perpustakaan Download (1MB)
	Text S_MKB_2103545_Chapter5.pdf Download (261kB)
	Text S_MKB_2103545_Appendix.pdf Restricted to Staf Perpustakaan Download (896kB)

Item Type:	Thesis (S1)
Additional Information:	https://scholar.google.com/citations?view_op=list_works&hl=id&user=4h1eJYEAAAAJ ID SINTA PEMBIMBING: Mahmudah Salwa Gianti: 6779018 Dewi Indriati Hadi Putri: 6720737
Uncontrolled Keywords:	CNN, Deepfake, Farneback, Optical Flow, Transformers CNN, Deepfake, Farneback, Optical Flow, Transformers.
Subjects:	T Technology > T Technology (General)
Divisions:	UPI Kampus Purwakarta > S1 Mekatronika dan Kecerdasan Buatan
Depositing User:	Muhammad Fariduddin Athar
Date Deposited:	09 Sep 2025 07:44
Last Modified:	09 Sep 2025 07:44
URI:	http://repository.upi.edu/id/eprint/138260

Actions (login required)

View Item