Analisis Perbandingan Kinerja Arsitektur CNN Dan Transformer dalam Klasifikasi Gambar AI-Generated dan Gambar Asli

The advancement of generative artificial intelligence has enabled the creation of increasingly realistic facial images that are difficult to distinguish from authentic images, thereby creating a need for automated classification methods capable of maintaining both accuracy and consistency in detection. This study compares the performance of two deep learning architecture groups, namely Convolutional Neural Networks (CNNs) and Transformers, for the binary classification of AI-generated images and real images. Four models were evaluated: ResNet50 and MobileNetV2 as CNN representatives, and Vision Transformer (ViT) and Data-efficient Image Transformer (DeiT) as Transformer representatives. The dataset was obtained from the 140k Real and Fake Faces dataset on Kaggle, with 28,000 images used in a balanced distribution: 20,000 images for training, 4,000 for validation, and 4,000 for testing. The experimental workflow followed the CRISP-DM framework, comprising business and data understanding, data preparation, modeling, evaluation, and deployment. Model performance was assessed using accuracy, precision, recall, F1-score, confusion matrix, and classification report. The results show that DeiT achieved the best performance, with an accuracy of 86.98%, precision of 86.81%, recall of 87.20%, and F1-score of 87.00%, followed by MobileNetV2, ViT, and ResNet50. In aggregate, the Transformer-based models achieved an average accuracy of 79.5%, which was 7% higher than that of the CNN-based models. The DeiT model was subsequently implemented in a Gradio-based prototype application capable of receiving user-uploaded images, performing inference, and displaying the predicted classification label along with its confidence score.