Hybrid Feature Extraction dengan XAI-Based Feature Refinement untuk Face Anti-Spoofing

Tsuroyya, Putri Ghaida (2026) Hybrid Feature Extraction dengan XAI-Based Feature Refinement untuk Face Anti-Spoofing. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221053-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (31MB) | Request a copy

Abstract

Sistem pengenalan wajah yang semakin luas digunakan rentan terhadap presentation attack seperti foto cetak, foto layar, dan wajah yang dibangkitkan kecerdasan buatan (Generate AI), sehingga dibutuhkan metode face anti-spoofing yang andal dan
interpretatif. Penelitian ini mengusulkan model klasifikasi wajah asli (real) dan wajah palsu (spoof) menggunakan data primer yang mencakup ketiga jenis serangan tersebut serta data sekunder dari dataset publik. Pra-pemrosesan meliputi deteksi dan cropping wajah menggunakan algoritma Viola–Jones, ekstraksi representasi tekstur Local Phase Quantization (LPQ), serta resize dan normalisasi. Model irancang berbasis arsitektur dual-stream EfficientNet-B0 pretrained ImageNet yang memproses citra RGB dan LPQ
secara paralel, masing-masing melalui komponen XAI-Based Feature Refinement berupa XAIPooling dan XAIDropout yang memanfaatkan nilai Shapley untuk memandu proses agregasi dan regulasi fitur secara adaptif. Representasi kedua stream digabungkan dan diklasifikasikan melalui tiga lapisan fully connected dengan fungsi aktivasi sigmoid. Model dievaluasi melalui 12 percobaan yang mengombinasikan dua skenario jumlah sampel latar belakang SHAP, tiga variasi dropout rate, dan dua interval pembaruan nilai Shapley. Hasil terbaik diperoleh pada konfigurasi dropout 0,45 dengan interval pembaruan nilai Shapley 10 epoch dan background samples 32, menghasilkan nilai APCER sebesar 0,0170, BPCER sebesar 0,0255, dan ACER sebesar 0,0213. Hasil ini
lebih baik dibandingkan tiga konfigurasi pembanding lain, yaitu LPQ Only, dual-stream RGB–LPQ dengan pooling dan dropout standar, serta RGB Only dengan XAIPooling dan XAIDropout, mengonfirmasi bahwa kombinasi dual-stream RGB–LPQ dengan XAIPooling dan XAIDropout saling melengkapi dalam meningkatkan kemampuan deteksi spoofing.
======================================================================================================================================
Face recognition systems, now widely deployed, remain vulnerable to presentation attacks such as printed photographs, screen-replay images, and artificially generated (Generative AI) faces, which calls for a reliable and interpretable face anti-spoofing method. This study proposes a classification model for genuine (real) and fake (spoof) faces using primary data covering all three types of attacks, combined with secondary data drawn from public datasets. Preprocessing consists of face detection and cropping using the Viola–Jones algorithm, texture representation extraction via Local Phase Quantization (LPQ), followed by resizing and normalization. The model is built on a dual-stream EfficientNet-B0 architecture pretrained on ImageNet, processing RGB and LPQ images in parallel, with each stream enhanced by an XAI-Based Feature Refinement component consisting of XAIPooling and XAIDropout, which leverage Shapley values to adaptively guide feature aggregation and regularization. The representations from both streams are fused and classified through three fully connected layers with a sigmoid
activation function. The model is evaluated across 12 experiments combining two SHAP background-sample scenarios, three dropout rate variations, and two Shapley value
update intervals. The best result is obtained with a dropout rate of 0.45 and a Shapley value update interval of 10 epochs, yielding an APCER of 0.0170, a BPCER of 0.0255,
and an ACER of 0.0213. These results outperform three other comparison configurations, namely LPQ Only, dual-stream RGB–LPQ with standard pooling and dropout, and RGB Only with XAIPooling and XAIDropout, confirming that the combination of the dual-stream RGB–LPQ architecture with XAIPooling and XAIDropout contributes in a complementary manner to improved spoofing detection performance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Face Anti-Spoofing, XAIPooling, XAIDropout, EfficientNet-B0, Local Phase Quantization, Explainable AI Face Anti-Spoofing, XAIPooling, XAIDropout, EfficientNet-B0, Local Phase Quantization, Explainable AI
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Putri Ghaida Tsuroyya
Date Deposited:	31 Jul 2026 06:40
Last Modified:	31 Jul 2026 06:40
URI:	http://repository.its.ac.id/id/eprint/140939

Actions (login required)

View Item