LEVERAGING HYBRID DEEP ENSEMBLES FOR CUTTING-EDGE IMAGE FORGERY DETECTION WITH ATTENTION-CNN, EFFICIENTNETB0, AND VISION TRANSFORMER

With the increasing realism of AI-generated images, distinguishing them from real ones has become a growing challenge. Previousstudies have used Convolutional Neural Networks (CNNs) for classifying synthetic and real images, but CNNs present severallimitations. They often fail to focus on critical regions, depend heavily on increasing depth—raising computational costs—and lackglobal context understanding due to localized receptive fields. To overcome these issues, this work introduces a hybrid deep learningapproach combining three advanced architectures: Attention-CNN with CBAM, EfficientNetB0, and Vision Transformer (ViT).CBAM enhances CNNs with spatial and channel attention, improving focus on key image features like textures and artifacts.EfficientNetB0 applies compound scaling to optimize network depth, width, and resolution for better performance with fewerresources. ViT captures global dependencies by treating images as patch sequences, enabling recognition of subtle, long-rangepatterns. The proposed ensemble was trained and evaluated on the CIFAKE dataset, which contains both real and AI-generatedimages. It achieved an accuracy of 96.81%, significantly outperforming standard CNNs. This study highlights the advantages ofcombining attention mechanisms and transformer-based models for more accurate and efficient synthetic image detection