Comparative Evaluation of CNN, Random Forest, and SVM for Deepfake Image Classification

Abstract

Deepfake image detection remains a challenging image classification problem because the effectiveness of a detection model can vary according to the classification approach and image representation used. This study comparatively evaluates three classification algorithms, namely Convolutional Neural Network (CNN), Random Forest, and Support Vector Machine (SVM), for distinguishing real facial images from deepfake images. The dataset consists of two classes, real and deepfake images, which undergo preprocessing through resizing to 32 × 32 pixels and pixel normalization from 0–255 to 0–1. The preprocessed dataset is divided into 80% training data and 20% testing data. CNN performs feature learning through convolutional layers, whereas Random Forest and SVM perform classification based on the prepared image feature representation. Model performance is evaluated using accuracy, precision, recall, and F1-score on the testing data. The experimental results show that CNN achieves the highest performance, with 91% accuracy, 93% precision, 88% recall, and 91% F1-score, followed by SVM with 77% accuracy and Random Forest with 75% accuracy. Thus, under the experimental conditions of this study, CNN provides better classification performance than the two conventional machine-learning algorithms. However, the result is limited to the dataset, 32 × 32 pixel input representation, and experimental configuration used in this study; therefore, the 91% accuracy should not be interpreted as evidence of general superiority across other datasets or deepfake generation techniques. These findings provide an empirical comparison of the three approaches and highlight the need for cross-dataset and higher-resolution evaluation in subsequent research.