Pengembangan Model Face Anti-Spoofing (FAS) Berbasis Convolutional Neural Network dan Deep Metric Learning

Rahmat, Ghulam Alvi (2026) Pengembangan Model Face Anti-Spoofing (FAS) Berbasis Convolutional Neural Network dan Deep Metric Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025232023-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Otentikasi biometrik wajah pada perangkat seluler krusial namun rentan terhadap serangan spoofing, diperparah oleh variabilitas kondisi penggunaan. Metode anti-spoofing konvensional dan klasifikasi Deep Learning (DL) standar memiliki keterbatasan signifikan dalam hal akurasi, keandalan, generalisasi terhadap serangan baru, dan efisiensi komputasi. Penelitian ini mengusulkan pendekatan Deep Metric Learning (DML) untuk mengembangkan sistem anti- spoofing wajah yang lebih efektif. DML bertujuan melatih model untuk memproyeksikan citra wajah ke dalam ruang laten (embedding space) sehingga representasi wajah autentik teraglomerasi rapat, sementara representasi wajah autentik dan palsu terpisah secara signifikan, umumnya melalui skema triplet loss. Penelitian ini akan fokus pada pengembangan model DML anti-spoofing yang ringan dan andal menggunakan arsitektur jaringan saraf efisien, dengan mengeksplorasi strategi sampling triplet dan dimensionalitas embedding optimal. Metodologi meliputi perancangan model DML, pengujian strategi sampling, dan analisis pengaruh dimensi embedding. Kinerja model DML akan dievaluasi secara komprehensif dan dibandingkan dengan metode baseline berdasarkan akurasi, keandalan terhadap variasi lingkungan dan serangan baru, serta efisiensi komputasi. Diharapkan penelitian ini menghasilkan sistem anti-spoofing yang akurat, andal, dan efisien.
=======================================================================================================================================
Facial biometric authentication on mobile devices is crucial but vulnerable to spoofing attacks, exacerbated by the variability of usage conditions. Conventional anti-spoofing methods and standard Deep Learning (DL) classifiers have significant limitations in terms of accuracy, robustness, generalization to new attacks, and computational efficiency. This paper proposes a Deep Metric Learning (DML) approach to develop a more effective facial anti-spoofing system. DML aims to train a model to project facial images into a latent space (embedding space) so that authentic facial representations are densely agglomerated, while authentic and fake facial representations are significantly separated, typically through a triplet loss scheme. This paper will focus on developing a lightweight and robust DML anti-spoofing model using an efficient neural network architecture, exploring optimal triplet sampling strategies and embedding dimensionality. The methodology includes designing a DML model, testing sampling strategies, analyzing the influence of embedding dimensions. The performance of the DML model will be comprehensively evaluated and compared with baseline methods based on accuracy, robustness to environmental variations and new attacks, and computational efficiency. It is expected that this research will produce an anti- spoofing system that is accurate, reliable, and efficient.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Arsitektur jaringan saraf, Deep Metric Learning (DML), Otentikasi biometrik wajah, Serangan spoofing, Deep Metric Learning (DML), Facial biometric authentication, Neural network architecture, Spoofing attacks
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Industrial Technology > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Ghulam Alvi Rahmat
Date Deposited:	27 Jan 2026 03:47
Last Modified:	27 Jan 2026 03:47
URI:	http://repository.its.ac.id/id/eprint/130416

Actions (login required)

View Item