Electronic Theses and Dissertation

Perkembangan teknologi generative model berbasis deep learning telah menghasilkan gambar sintetis berkualitas tinggi yang semakin sulit dibedakan dari gambar asli, menimbulkan tantangan serius dalam verifikasi keaslian konten visual digital. Penelitian ini bertujuan mengembangkan dan membandingkan model klasifikasi berbasis ConvNeXt dan ResNet-50 untuk mendeteksi gambar hasil generasi kecerdasan buatan (AI Generated) dan gambar asli menggunakan dataset CIFAKE yang terdiri dari 120.000 citra turunan CIFAR-10 dengan distribusi kelas seimbang. Kedua model dilatih menggunakan strategi progressive fine-tuning tiga fase dengan bobot pretrained ImageNet, optimizer AdamW, dan augmentasi data komprehensif. Hasil penelitian menunjukkan bahwa ConvNeXt mencapai test accuracy 98,70% dengan macro F1-Score 0,9870, melampaui ResNet-50 yang mencapai test accuracy 97,95% dengan macro F1-Score 0,9795. Dari sisi efisiensi komputasi, ResNet-50 lebih unggul dengan waktu inferensi 3,2780 ms per citra dan throughput 305,1 img/s dibandingkan ConvNeXt sebesar 4,6167 ms dan 216,6 img/s. Kesimpulan penelitian ini adalah ConvNeXt merupakan model yang paling optimal dengan keseimbangan terbaik antara performa klasifikasi dan efisiensi komputasi, dan selanjutnya diimplementasikan ke dalam prototipe aplikasi web berbasis Flask sebagai sistem deteksi gambar AI Generated secara otomatis.

The rapid advancement of deep learning-based generative models has enabled the synthesis of highly realistic images that are increasingly indistinguishable from authentic photographs, posing significant challenges to the integrity of digital visual content verification. This study aims to develop and compare two transfer learning-based classification models, ConvNeXt and ResNet-50, for detecting AI-generated images against real images using the CIFAKE dataset, which comprises 120,000 images derived from CIFAR-10 with a balanced class distribution. Both models were trained using a three-phase progressive fine-tuning strategy with ImageNet pretrained weights, the AdamW optimizer, and comprehensive data augmentation. Experimental results demonstrate that ConvNeXt achieves a superior test accuracy of 98.70% with a macro F1-Score of 0.9870, outperforming ResNet-50, which attains a test accuracy of 97.95% with a macro F1-Score of 0.9795. In terms of computational efficiency, however, ResNet-50 proves more favorable, delivering an inference time of 3.2780 ms per image and a throughput of 305.1 img/s, compared to ConvNeXt at 4.6167 ms and 216.6 img/s, respectively. Overall, ConvNeXt is established as the most optimal architecture, offering the best balance between classification performance and computational efficiency, and is subsequently deployed in a Flask-based web application prototype as an automated AI-generated image detection system.