OPTIMASI HYPERPARAMETER PADA ARSITEKTUR MODEL CNN LIGHTWEIGHT UNTUK DETEKSI SPOOFING WAJAH MENGGUNAKAN OPTUNA

Prihanto Dwi Rahmanto, NIM.: 22206051021 (2025) OPTIMASI HYPERPARAMETER PADA ARSITEKTUR MODEL CNN LIGHTWEIGHT UNTUK DETEKSI SPOOFING WAJAH MENGGUNAKAN OPTUNA. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (OPTIMASI HYPERPARAMETER PADA ARSITEKTUR MODEL CNN LIGHTWEIGHT UNTUK DETEKSI SPOOFING WAJAH MENGGUNAKAN OPTUNA) 22206051021_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (6MB) | Preview
	Text (OPTIMASI HYPERPARAMETER PADA ARSITEKTUR MODEL CNN LIGHTWEIGHT UNTUK DETEKSI SPOOFING WAJAH MENGGUNAKAN OPTUNA) 22206051021_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (14MB) | Request a copy

Abstract

Traditionally, hyperparameter tuning relies on grid search or random search, but these methods suffer from inefficiencies in exploring large hyperparameter spaces and often demand extensive computational time. To address these limitations, researchers increasingly use Bayesian optimization, which builds a surrogate model from past trials to guide the selection of promising hyperparameters. In this study, we applied Bayesian optimization using the Optuna framework - chosen for its flexible "define-by-run" interface - to improve face spoofing detection with lightweight CNNs: LMFRNet, MobileNetV3 Small, ShuffleNetV2, and ResNet50, on the CelebA-Spoof dataset. The preprocessing pipeline included MTCNN-based face detection and alignment, followed by data cleaning, CLAHE for brightness enhancement, and undersampling for class balance. Transfer learning was then applied to train the CNN models. Bayesian optimization yielded these optimal hyperparameters: for ShuffleNetV2, ResNet50, and MobileNetV3 Small, a learning rate of 0.001, batch size of 64, six epochs, and SGD optimizer; for LMFRNet, a learning rate of 0.01, batch size of 32, eight epochs, and SGD. Among these, ShuffleNetV2 achieved the best performance, reaching 98% accuracy.

Share this knowledge with your friends :

Actions (login required)

View Item