ENHANCING TRUST: EXPLAINABLE AI FOR IDENTIFYING GENERATED IMAGES

Generative models have achieved rather rapid progress, which has simplified distinguishing between real and fake photos. The existing detection techniques are typically not very practical or provide unclear explanations, making them less applicable in practice. This paper offers a DL-based model of image classification with explainable identification to address this problem. The system works using a tagged set of actual and AI-generated photos that have been changed in size, normalized, and augmented by flipping, rotating, and changing the color. Some of the models we use and test include a custom CNN, a longer CNN with dropout, and a pretrained Xception network. The Flask-based web app uses the best-performing model to make predictions in real time. Also, a Grad-CAM mechanism is included in to make visual explanations that show crucial areas that affect the model’s choice. The accuracy of the Xception model is 99.92% that indicates that it is superior to other models in experiments. The proposed method is an effective combination of high classification and interpretability, making it suitable to real-world synthetic image detection tasks. “