Implementasi Pendekatan Ensemble Deep Learning untuk Klasifikasi Citra Ilustrasi Asli dan Hasil Generative AI

The rapid development of Generative Artificial Intelligence (AI) has increasingly enabled the creation of digital illustrations with high visual similarity to human-created artworks, making identification more difficult and raising challenges related to originality and copyright protection. However, research on Generative AI image detection has largely focused on photographic images or real-world objects, while the classification of digital illustrations with artistic characteristics and high visual similarity remains limited. This study aims to implement and evaluate an automated classification method to distinguish original illustrations from Generative AI-generated illustrations. The proposed approach employs Ensemble Deep Learning with a Soft Voting technique that combines MobileNetV2, ResNet50, and DenseNet121. The modeling process follows the CRISP-DM methodology and is evaluated using 5-Fold Stratified Cross Validation. Internal validation achieved an accuracy of 96.00% and an AUC of 99.06%. In external testing, accuracy reached 87.60% with an AUC of 95.09%, while individual-sample testing achieved an accuracy of 80.00%. The decrease in accuracy on external data indicates challenges in generalizing the model to data with characteristics different from those in the training and validation sets. This study contributes by applying and evaluating a Soft Voting Ensemble for classifying anime- and manga-style digital illustrations through internal validation, external testing, and individual-sample testing. The results indicate that this approach has the potential to distinguish original illustrations from Generative AI-generated illustrations, although further improvements in the model's generalization capability are needed.