Deepfake Image Detection Using HOG-SVM Face Extraction and EfficientNetB5-Based Deep Learning Framework

Abstract

Deep learning has made significant strides recently and has produced amazing lifelike-looking synthetic images called deepfakes, making it very difficult to differentiate between real images and fake images. The abuse of deepfake technology can lead to issues like misinformation, identity fraud, and a loss of confidence in digital media. This paper presents a framework that detects deepfake images by combining traditional picture processing techniques and deep learning methods. At first, face images are detected and extracted from the original image by calculating HOG features and classifying them with SVM (Support Vector Machines). The collected face images are then processed using a pre-trained EfficientNetB5 network trained on non-manipulated images by way of transfer learning to identify visual differences between original images and deepfakes produced by manipulating images. Finally, the system can classify the images as either real or deepfake using binary classification methods. The performance of the proposed framework is evaluated using various performance evaluation metrics, such as accuracy, precision, recall, F1-score, and ROC-AUC, confusion matrixes. The results show that the model obtained 97.7% accuracy, 97.3% precision, 97.5% recall, and 0.975 ROC-AUC score; thus, indicating that the model can successfully detect deepfakes with good accuracy and has good generalization capability for unseen datasets. Overall, the proposed framework offers a very viable, scalable solution to real-world digital forensics and deepfake detection applications.