Enhanced CNN Architecture with Residual Blocks and Regularization for AI-Generated Image Detection
Jayanti Rout, Minati Mishra · 2025
The growing popularity of social networks has led to an unprecedented surge in the number of digital images shared daily. As a result, ensuring the authenticity of these images has become a pressing concern, especially with the rise of advanced manipulation techniques. Such fake images are an issue for citizens around the world as they not only spread fake information, but also harm reputations and cause financial harm as well. Among artificial intelligence (AI)-generated image domains, Generative Adversarial Networks (GANs), deepfakes, and text-to-image diffusion models are prominent examples. They use powerful deep learning models to create highly convincing fake content. This rapid advancement has presented significant challenges for conventional detection strategies, causing them to struggle to keep up. A deep learning-based method is proposed to detect AI-generated images using enhanced convolutional neural networks (CNNs). The CNN model is designed to classify images as authentic or fake. The model is trained and evaluated on the 140k Real and Fake Faces (RFF) dataset. We evaluated the model using various performance metrics and discovered that it outperforms similar models. The model achieves a strong accuracy of 98.125% in the test set, demonstrating its potential effectiveness in digital image forgery detection.