Klasifikasi Karya Seni Asli dan Sintesis AI untuk Deteksi Plagiarisme Visual Menggunakan CNN dan Data-Efficient Image Transformer

Tjias, Richie Leonard Tjias (2026) Klasifikasi Karya Seni Asli dan Sintesis AI untuk Deteksi Plagiarisme Visual Menggunakan CNN dan Data-Efficient Image Transformer. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221008-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (3MB)

Abstract

Pesatnya perkembangan kecerdasan buatan (AI) dalam menghasilkan karya seni visual memicu masalah mengenai orisinalitas dan potensi plagiarisme visual bagi seniman. Penelitian ini bertujuan untuk mengklasifikasikan karya seni asli buatan manusia dan sintesis AI menggunakan arsitektur Data-efficient image Transformer (DeiT) dengan membandingkan performanya terhadap model Convolutional Neural Network (CNN). Metode yang diterapkan adalah Knowledge Distillation, di mana model CNN bertindak sebagai guru untuk membantu dalam melatih model DeiT melalui mekanisme hard-label distillation. Data yang digunakan berjumlah 21.642 gambar yang dibagi menjadi dataset pelatihan, validasi, dan pengujian. Pra-pemrosesan data mencakup resize ke 224×224 piksel, normalisasi, serta augmentasi MixUp dan CutMix untuk meningkatkan peforma model. Hasil eksperimen menunjukkan bahwa model DeiT berhasil mencapai akurasi uji tertinggi sebesar 0,9635 (96,35%) dan F1-Score stabil sebesar 0,9635 untuk kedua kelas, mengungguli model CNN rangkaian sendiri yang hanya memperoleh akurasi uji sebesar 0,8873 (88,73%) dan nilai F1-Score sekitar 0,88 untuk kedua kelas. Penggunaan distillation token dalam model DeiT terbukti efektif dalam mengenali anomali gambar yang dihasilkan AI. Namun, pada proses inferensi ditemukan adanya batasan pada model dalam mengklasifikasikan gambar yang dihasilkan AI terbaru. Dapa disimpulkan, pendekatan DeiT dengan strategi Knowledge Distillation bisa menjadi solusi yang akurat dan robust dalam memvalidasi keaslian karya seni di era digital.
=================================================================================================================================
The rapid development of artificial intelligence (AI) in generating visual artwork has raised significant concerns regarding originality and the potential for visual plagiarism among artists. This study aims to classify human-made original artwork and AI-synthesized art using the Data-efficient image Transformer (DeiT) architecture, comparing its performance against a Convolutional Neural Network (CNN) model. The Knowledge Distillation method is applied, where the CNN model acts as a teacher to assist in training the DeiT model through a hard-label distillation mechanism. The dataset comprises 21,642 images, split into training, validation, and testing subsets. Data preprocessing includes resizing images to 224×224 pixels, normalization, as well as MixUp and CutMix augmentations to enhance model performance. Experimental results demonstrate that the DeiT model achieves the highest test accuracy of 0,9635 (96.35%) and a stable F1-Score of 0,9635 for both classes, outperforming the custom-built CNN model, which only achieves a test accuracy of 0,8873 (88.73%) and an F1-Score of approximately 0,88 for both classes. The integration of the distillation token within the DeiT model proves effective in detecting anomalies in AI-generated images. However, during the inference process, limitations were identified in classifying images generated by state-of-the-art AI models. In conclusion, the DeiT approach combined with a Knowledge Distillation strategy provides an accurate and robust solution for validating the authenticity of visual artwork in the digital era.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Kecerdasan Buatan, MixUp, CutMix, Piksel, plagiarisme, Model, Generative AI, Token, Knowledge Distillation
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Richie Leonard Tjias
Date Deposited:	05 Aug 2026 03:49
Last Modified:	05 Aug 2026 03:49
URI:	http://repository.its.ac.id/id/eprint/143906

Actions (login required)

View Item