ANALISIS PERBANDINGAN METODE FUSI MULTIMODAL UNTUK DETEKSI PEMALSUAN WAJAH: CONCATENATION, ADDITION, DAN CROSS-ATTENTION

SANI, SAYANG (2025) ANALISIS PERBANDINGAN METODE FUSI MULTIMODAL UNTUK DETEKSI PEMALSUAN WAJAH: CONCATENATION, ADDITION, DAN CROSS-ATTENTION. Skripsi thesis, UPN Veteran Yogyakarta.

	Text ABSTRAK.pdf Download (179kB)
	Text COVER.pdf Download (344kB)
	Text DAFTAR ISI.pdf Download (189kB)
	Text DAFTAR PUSTAKA.pdf Download (111kB)
	Text HALAMAN PENGESAHAN PEMBIMBING.pdf Download (548kB)
	Text HALAMAN PENGESAHAN PENGUJI.pdf Download (618kB)
	Text NASKAH_TA_123210044_SAYANGSANI.pdf Restricted to Repository staff only Download (12MB)

Abstract

1 ABSTRAK
Teknologi pengenalan wajah (face recognition) banyak digunakan namun rentan
terhadap serangan pemalsuan (spoofing attack), terutama print attack yang mudah
dieksekusi. Untuk mengatasi keterbatasan sistem berbasis RGB tunggal, pendekatan multi
modalitas yang menggabungkan data RGB, Depth, dan Inframerah (IR) menjadi solusi
efektif karena setiap modalitas menawarkan petunjuk unik untuk membedakan wajah asli
dan palsu. Namun, keberhasilan pendekatan ini sangat bergantung pada strategi fusi fitur
(feature fusion) yang digunakan, dan banyak penelitian sebelumnya membandingkan
strategi ini menggunakan arsitektur yang berbeda-beda sehingga perbandingan menjadi
tidak adil. Penelitian ini bertujuan untuk menganalisis secara komparatif efektivitas tiga
metode fusi fitur Concatenation, Addition, dan Cross-Attention dalam kerangka kerja yang
terkontrol. Untuk memastikan perbandingan yang adil, ketiga strategi diimplementasikan
pada arsitektur backbone ResNet-50 yang seragam dan diuji pada dataset multi-modal
CASIA-SURF. Hasil penelitian menunjukkan bahwa metode Addition Fusion mencapai
kinerja terbaik dengan Average Classification Error Rate (ACER) 0.68%, akurasi 99.29%,
dan F1-Score 99.49%. Kinerja ini secara signifikan mengungguli Cross-Attention Fusion
(ACER 0.73%) dan Concatenation Fusion (ACER 1.37%).
Temuan utama penelitian ini adalah superioritas metode Addition Fusion yang secara
arsitektural lebih sederhana dibandingkan Cross-Attention yang lebih kompleks, yang
mengindikasikan bahwa agregasi fitur komplementer secara langsung lebih robust dan
efektif untuk tugas deteksi print attack. Penelitian ini menyajikan sebuah benchmark
kuantitatif yang terkontrol dan menantang asumsi bahwa arsitektur yang lebih kompleks
selalu memberikan hasil yang lebih baik.
Kata Kunci: Face Anti-Spoofing, Deteksi Pemalsuan Wajah, Fusi Multi-modal, Fusi Fitur,
Concatenation, Addition, Cross-Attention, ResNet-50, CASIA-SURF.
vi
2 ABSTRACT
Face recognition technology is widely used but is vulnerable to spoofing attacks ,
particularly print attacks, which are easy to execute. To overcome the limitations of single
modality RGB-based systems , a multi-modal approach combining RGB, Depth, and Infrared
(IR) data has emerged as an effective solution , as each modality offers unique cues to
distinguish between real and fake faces. However, the success of this approach is highly
dependent on the feature fusion strategy employed. Many previous studies have compared
these strategies using different underlying architectures, making fair comparisons difficult.
This research aims to comparatively analyze the effectiveness of three feature fusion methods
Concatenation, Addition, and Cross-Attention—within a controlled framework. To ensure a
fair comparison, all three strategies were implemented on a uniform ResNet-50 backbone
architecture and tested on the CASIA-SURF multi-modal dataset. The results show that the
Addition Fusion method achieved the best performance, with an Average Classification
Error Rate (ACER) of 0.68%, an accuracy of 99.29%, and an F1-Score of 99.49%. This
performance significantly surpassed both Cross-Attention Fusion (ACER 0.73%) and the
baseline Concatenation Fusion (ACER 1.37%).
The main finding of this study is the superiority of the architecturally simpler
Addition method over the more complex Cross-Attention , indicating that the direct
aggregation of complementary features is more robust and effective for the task of print
attack detection. This research provides a controlled, quantitative benchmark and challenges
the assumption that more complex architectures always yield better results.
Keyword: Face Anti-Spoofing, Face Spoofing Detection, Multi-modal Fusion, Feature
Fusion, Concatenation, Addition, Cross-Attention, ResNet-50, CASIA-SURF.
vii

Item Type:	Tugas Akhir (Skripsi)
Uncontrolled Keywords:	Face Anti-Spoofing, Deteksi Pemalsuan Wajah, Fusi Multi-modal, Fusi Fitur, Concatenation, Addition, Cross-Attention, ResNet-50, CASIA-SURF.
Subjek:	Z Bibliography. Library Science. Information Resources > ZA Information resources
Divisions:	Fakultas Teknik Industri > (S1) Informatika
Depositing User:	Eko Yuli
Date Deposited:	10 Oct 2025 02:06
Last Modified:	10 Oct 2025 02:06
URI:	http://eprints.upnyk.ac.id/id/eprint/44213

Actions (login required)

View Item