Implementing a Multi-Model Machine-Learning Approach to Face Detection and Feature Extraction for Still-Image Deepfake Detection: A Simulation-Based Evaluation

Abstract

Background. Detecting a manipulated face in a still image depends on two choices made before any classifier is trained: which face-detection front-end locates and crops the face, and which feature-extraction representation that crop is turned into. Rahman et al. investigated face-detection and feature-extraction techniques for deepfake detection using a multi-model machine-learning approach, comparing how different detector and feature choices affect downstream classification. Methods. This study implements that investigation directly: two face-detection front-ends (a Viola-Jones-style Haar-cascade proxy and an MTCNN-style landmark-aligned proxy, differing in how much alignment jitter they introduce), three feature-extraction representations (local-binary-pattern texture features, histogram-of-oriented-gradients features, and Xception-style deep-embedding features), and three classifiers (a support-vector machine, a Random Forest, and a small multi-layer perceptron) are combined into an eighteen-cell grid and evaluated, over five seeds, on a synthetic but literature-informed feature-vector-level face-crop dataset, since real video-level deepfake corpora and full convolutional-network training are infeasible in this evaluation's computing environment. Results. Feature-extraction choice dominated the grid: deep-embedding features reached a mean F1 of 0.897-0.925 depending on detector, against 0.732-0.803 for gradient (HOG) features and 0.777-0.811 for texture (LBP) features. Detector choice mattered consistently but more modestly, with the landmark-aligned detector outperforming the Haar-cascade-style detector on every feature type (F1 advantage of 0.025-0.028). Classifier choice mattered least of the three factors at the best detector-feature combination, with Random Forest, SVM, and the MLP within 1.2 F1 points of one another (0.919-0.931). The single strongest pipeline combined the landmark-aligned detector, deep-embedding features, and Random Forest (F1 0.931, AUC 0.979). Conclusion. Implementing Rahman et al.'s multi-model investigation end to end on this evaluation's synthetic face-crop data shows that feature-extraction quality, more than either face-detection front-end or final classifier choice, is the dominant lever for still-image deepfake-detection accuracy, a finding with direct implications for where a practitioner with a fixed computing budget should invest first.