Enhancing Fake Image Detection with Ensembled Convolutional Neural Networks

Fake image detection has emerged as a vital task for the Generative AI era due to the fast evolution in generations of models that have made highly realistic synthetic images possible. In this paper, we formulate an ensemble-based Convolutional Neural Network (CNN) to enhance fake image detection accuracy. Our methodology includes the training of five CNN models on separate datasets consisting of real and artificially created images found in different public datasets. The artificially created images in the datasets are produced using the latest models that include StyleGAN2, StyleGAN3, Diffusion GAN, Taming Transformer and Gansformer. The outputs of the five CNN models are fused using a stacking ensemble process in which several different classifiers such as Random Forest, Gradient Boosting, AdaBoost, Support Vector Machine, Multi-Layer Perceptron and Logistic Regression are utilized to boost the final classification performance. The ultimate test on unseen data reveals an increase in the classification performance as our approach exhibits a high accuracy rate of more than 90%. Comparison of the performance of different classifiers utilized in the stacking ensemble and accuracy metrics such as precision, recall and F1-score reveals a complete insight about the performance of the proposed approach. These results indicate that the use of ensemble-based deep learning approaches makes fake image detection systems strongly robust in nature and even more applicable in real-world settings.