MEXFIC: A meta ensemble eXplainable approach for AI-synthesized fake image classification

Abstract

In the evolving landscape of artificial intelligence (AI), differentiating between authentic and artificially generated images poses a significant challenge, primarily due to the rapidly enhancing quality of AI-generated images. This paper systematically evaluates state-of-the-art classification models to distinguish authentic images from those synthetically produced using the CIFAKE dataset. We introduce FakeGPT and PFake, two new test datasets featuring genuine and AI-generated synthetic images with specific keywords paralleling the generation of the CIFAKE dataset. We use the transfer learning technique to train the state-of-the-art classification models on the CIFAKE training set, followed by rigorous evaluation against the CIFAKE, FakeGPT, and PFake test datasets. Further, we explore ensemble approaches, including stacking, voting, bagging, and meta-ensemble learning. The culmination of our extensive research efforts is the Meta Ensemble eXplainable Fake Image Classifier (MEXFIC), which stands out with a notable accuracy of 94% and 96.61% against the Stable Diffusion generated CIFAKE and PFake datasets, respectively. This is a significant improvement over the ConvNextLarge model, achieving the highest accuracy of 92.54% among the state-of-the-art models. Our study showcases the competitive edge of MEXFIC that highlights the necessity for more robust models capable of identifying AI-synthesized images, as evidenced by the performance on the challenging FakeGPT dataset.

Keywords

• Fake image classification
• AI-synthesized image classification
• AI-generated image classification
• Image authenticity verification
• Smart surveillance
• CIFAKE