Robust Metadata Forensics to Improve Detection and Analysis of Deepfake Images
Neil Royan, Cutifa Safitri, Tjong Wan Sen · 2024
The rapid advancement of deepfake technology presents significant challenges to digital forensics, particularly in accurately detecting and analyzing manipulated image metadata. Traditional forensic methods often struggle to keep pace with these sophisticated manipulations, undermining the reliability of digital evidence essential for legal and security purposes. This study is concerned with examining machine learning models’ efficiency in detecting real and fake images using metadata features relevant to digital forensic analysis. The research compares different models, such as Stacking Model, CatBoost, LightGBM, XBOOST, Gradient Boosting, Decision Tree, SVM, KNN, LR, Random Forest, CNN, VGG16, Augmented model, and Ensemble Learning Model. The result indicators show that the best one was Gradient Boosting, which achieved the highest accuracy of 86.02% and acceptable precision and recall in detecting both kinds of images. Apart from this, KNN also showed a productive result, with better precision in detecting real images; and Random Forest was the best in fake image detection.