A Comparative Benchmarking Study of Classical Machine Learning and Deep Learning Methods for Image-Based Deepfake Detection

Deepfake media provides a very genuine threat to digital trust, enabling misinformation, fraud, and impersonation of identity. This paper examines a comparison benchmark analysis between classical machine learning (ML) approaches using Support Vector Machines, Decision Tree, Random Forest, and Gradient Boosting as well as deep learning (DL) methods based on Convolutional Neural Networks (CNNs) for image-based deepfake detection. Although FaceForensics++ and Celeb-DF are video datasets, a fixed number of face frames were extracted and processed to construct a balanced dataset of 3930 images. Experimental evaluation shows that classical ML models achieve accuracies ranging from 84% to 91%, with gradient boosting performing the best among these, while the CNN achieves the highest accuracy of 93%. Though it has shown increased precision in its capability to detect, traditional models also provide an edge in terms of computational efficiency and interpretability. The significance of this particular study has been to bring forward inherent trade-offs and provide selection guidance for deepfake models.