OPTIMALISASI ARSITEKTUR CNN UNTUK DETEKSI PEMALSUAN WAJAH PADA SISTEM PRESENSI 

OPTIMIZING CNN ARCHITECTURE FOR FACE SPOOFING DETECTION IN ATTENDANCE SYSTEMS

Penggunaan sistem presensi berbasis pengenalan wajah semakin luas karena efisien dan praktis, namun sistem ini rentan terhadap serangan pemalsuan wajah (face spoofing), seperti penggunaan foto atau video. Oleh karena itu, diperlukan model deteksi face spoofing yang memiliki kemampuan generalisasi yang baik. Penelitian ini bertujuan untuk membandingkan performa arsitektur Convolutional Neural Network (CNN) MobileNetV2, EfficientNetB0, dan ResNet50 dalam mendeteksi serangan pemalsuan wajah melalui pengujian lintas dataset (cross-dataset evaluation), serta mengimplementasikan arsitektur CNN yang paling stabil pada sistem presensi berbasis wajah. Metode penelitian yang digunakan adalah penelitian eksperimental dengan desain perbandingan model menggunakan beberapa dataset face anti-spoofing. Data citra diperoleh melalui ekstraksi frame video setelah pembagian data pelatihan, validasi, dan pengujian untuk mencegah kebocoran data. Evaluasi performa dilakukan menggunakan metrik akurasi dan drop rate. Hasil penelitian menunjukkan bahwa EfficientNetB0 memiliki kestabilan performa terbaik dan selanjutnya dioptimalisasi untuk diimplementasikan pada sistem presensi berbasis wajah yang dikembangkan secara tidak real-time menggunakan Streamlit.

Kata kunci: face anti-spoofing, CNN, cross-dataset evaluation, EfficientNetB0, sistem presensi wajah

The use of face recognition–based attendance systems has become increasingly widespread due to their efficiency and practicality; however, such systems are vulnerable to face spoofing attacks, such as the use of photos or videos. Therefore, a face spoofing detection model with strong generalization capability is required. This study aims to compare the performance of Convolutional Neural Network (CNN) architectures—MobileNetV2, EfficientNetB0, and ResNet50—in detecting face spoofing attacks through cross-dataset evaluation, as well as to implement the most stable CNN architecture in a face recognition–based attendance system. This research employs an experimental method with a model comparison design using several face anti-spoofing datasets. Image data are obtained through video frame extraction after splitting the data into training, validation, and testing sets to prevent data leakage. Model performance is evaluated using accuracy and drop rate metrics. The results indicate that EfficientNetB0 demonstrates the best performance stability and is subsequently optimized for implementation in a non-real-time face recognition–based attendance system developed using Streamlit.

Keywords: face anti-spoofing, CNN, cross-dataset evaluation, EfficientNetB0, face recognition–based attendance system