ANALISIS PERFORMA CONVOLUTIONAL NEURAL NETWORK DENGAN HYPERPARAMETER TUNING DALAM MENDETEKSI GAMBAR DEEPFAKE

Abstract

This research analyzes the performance of Convolutional Neural Network (CNN) in detecting deepfake images with a focus on hyperparameter tuning. The study consists of two classes: fake images and real images, with each class containing 5000 data samples. Hyperparameter tuning is conducted using the Keras-tuner library, a framework used for automatic hyperparameter tuning on models built with Keras, eliminating the need for manual trial and error tuning. The hyperparameter search strategy employed is random search. The results of the study indicate that hyperparameter tuning significantly improves the model's detection accuracy. Various experiments were conducted to evaluate the impact of hyperparameter settings, such as the number and size of filters, learning rate, and optimizer. Analysis of different optimizers showed significant variations in performance, with Adam Optimizer achieving the highest accuracy of 83% using a combination of 32 filters sized 3x3 in the first layer and 128 filters sized 5x5 in the second layer. RMSProp and AdamW each achieved 82% accuracy, SGD Optimizer achieved 75% accuracy, while Adadelta Optimizer achieved 71% accuracy. The findings of this study affirm that the selection of optimizer and appropriate hyperparameter settings have a significant impact on the model's performance in detecting patterns in the data. This study also emphasizes the importance of optimizing filters and sizes in each layer to enhance model accuracy.