CIFAKE: Explainable Image Classification and Identification of AIGenerated Synthetic Images

Abstract

Humans now find it more difficult to distinguish between artificial intelligence (AI)- generated and real-world pictures due to recent developments in synthetic data synthesis. By offering a technique to improve the identification of AIgenerated pictures using computer vision algorithms, this study meets the crucial demand for data authenticity and trustworthiness. The study creates a set of pictures for comparison with actual photos using a synthetic dataset that was constructed using latent diffusion and patterned after the CIFAR-10 dataset. The classification task is presented as a binary problem that entails differentiating between images created by artificial intelligence and those that are real. For this, a CNN2D model—which has 32 neurons and produces the best results—is used. This structure consists of two convolutional layers, two MaxPooling2D layers, and two dense layers. Furthermore, the Grad-CAM (Gradient Class Activation Mapping) method identifies elements that help CNN differentiate between authentic and fraudulent pictures. In contrast to the suggested CNN2D's 94.98% accuracy, a modified version that was optimized without the addition of extra layers like dropout or global average pooling obtained 95.94%. Accuracy, precision, recall, F1-score, and a confusion matrix are used to assess the performance.