Analisis Deteksi Manipulasi Gambar Menggunakan Error Level Analysis dan Deep Learning di SD Padedewatu

Abstract

Digital images are increasingly used as documentation in educational environments, while image-editing tools make manipulation increasingly accessible. This study analyzes a workflow that combines Error Level Analysis (ELA) and a Convolutional Neural Network (CNN) for preliminary detection of manipulated images using photographs from SD Padedewatu as a case study. The dataset initially consisted of 34 original photographs. Because the supplied photographs did not contain verified manipulation labels, controlled manipulated variants were generated from each source image using copy-move, splicing, and object-erasure operations. The resulting labeled dataset contained 136 images, consisting of 34 authentic and 102 manipulated images. ELA was applied by recompressing each image and visualizing the resulting compression-error differences. A CNN classifier was then trained on ELA representations. To reduce source-image leakage, 24 source-image groups were used for training and 10 source-image groups were reserved for testing, producing 96 training images and 40 testing images. The model achieved 75% test Accuracy, 75% precision, 100% recall, and an F1-score of 85.71%. However, the confusion matrix showed that all test images were predicted as manipulated, including the 10 authentic images. Therefore, the Accuracy value should not be interpreted as evidence of balanced or robust forgery detection. The experiment provides a reproducible local ELA-CNN workflow and identifies dataset balance, sample size, and external validation as important priorities for subsequent research.