Performance Analysis of the LBP–SVM Model in Deepfake Image Detection: A Case Study on the FaceForensics++ Dataset and External Validation on Indonesian Politicians’ Faces

Abstract

DeepFake images synthetic face images generated by deep neural networks (GANs) can seriously undermine media authenticity and public trust This study analyzes deepfake image detection based on handcrafted features using Local Binary Pattern (LBP) and Support Vector Machine (SVM). The primary dataset consists of 2,000 images from FaceForensics++, comprising 1,000 authentic and 1,000 manipulated images, with a 70:30 stratified train-test split. Each image undergoes preprocessing steps including grayscale conversion, intensity normalization, and resizing to 128×128 pixels before feature extraction using the uniform LBP operator (P=8, R=1). An SVM with a Radial Basis Function (RBF) kernel is employed as the baseline classifier, achieving an accuracy of 0.63, a macro-F1 score of 0.63, and a ROC–AUC of 0.65, demonstrating higher sensitivity to the fake class. To assess cross-domain generalization, an external validation was conducted using 10 images of Indonesian politicians (6 real, 4 fake). The model correctly classified 6 out of 10 images but showed a bias toward the real class due to differences in texture and lighting. These results suggest that the LBP–SVM approach remains relevant for lightweight texture-based deepfake detection, though it is not yet optimal for real-world domain variations. Future research is recommended to explore LBP combined with HOG or deep feature embeddings to improve accuracy and model robustness in practical scenarios.