Detecting AI Generated Fake Images Using Deep Learning

Abstract

The realism in AI-generated images is posing a significant problem in verifying the authenticity of digital content. In this context, this project focuses on developing a deep learningbased system for detecting fake images created using AI through an ensemble of convolutional neural networks, such as DenseNet121, ResNet50, and EfficientNet. An ensemblebased approach is proposed to increase the reliability of classification through soft voting using different architectures. The proposed system uses transfer learning, in which pre-trained models are fine-tuned using a dataset with real and synthetic images. Through this approach, the system can efficiently extract features from images and detect subtle differences in texture, structure, and pixel value distribution, which is challenging to do through manual analysis. Using this approach, the proposed system can effectively differentiate between real and AI-generated images with higher precision using hierarchical feature learning .For better interpretability, Grad-CAM is also proposed to visualize images to understand which parts of the images are affecting the decision made by the proposed system. Through this approach, it is possible to validate the decision made by the proposed system. Experimental results have shown that the ensemble-based approach outperforms individual CNNs in accuracy, generalization, and robustness, which is useful for detecting fake.