A robust ensemble model for Deepfake detection of GAN-generated images on social media

Abstract The emergence of deepfake images created by GANs models for malevolent purposes presents a serious risk to society as well as great challenge to digital security and trust. Leveraging the power of ensembles and combining machine and deep learning approaches, this paper presents VOTSTACK, an innovative ensemble model designed to combat the proliferation of deepfake images on social media. VOTSTACK utilizes a blended approach that combines Voting and Stacking ensemble techniques. It leverages the collective intelligence of three different classifiers—Decision Tree, Logistic Regression, and SVM—executing a hybrid feature selection method with Principal Component Analysis (PCA) conditioning as the preprocessing framework. It refines the features using iterative feature resolution with cross-validation (RFECV) method. This model operates through a two-phase architecture, with the first phase consolidating results using a voting ensemble and the second phase aggregating collective knowledge into a final decision using a stacking ensemble. A majority vote method is used in the first phase to aggregate predictions from the three base classifiers (Decision Tree, Logistic Regression, and SVM). Utilizing strength of each classifier and results of voting method as meta classifier, a stacking ensemble further refines these predictions in the second phase. The effectiveness and reliability of this approach is validated on a substantial dataset known as Real and Fake Images reliability. The proposed model outperforms conventional methods, achieving an impressive accuracy rate of 91.6%, a high precision score of 90.3%, a substantial recall of 89.8%, and an outstanding F1-Score of 90%.