CIFAKE: Precision Image Classification and Explainable AI for Detecting Synthetic Generations

Authors- D. Chakra Satya Tulasi, Singuluri Annapurna, Pasumarthi Samyuktha, Busi Srikanth, Golla Sneha Ratna, Gampala Lakshmi Shiva Teja

Abstract-The rapid advancements in synthetic data generation have resulted in AI-generated images that are nearly indistinguishable from real photographs, posing challenges in data authenticity and reliability. This study aims to enhance the detection of AI-generated images using computer vision techniques. A synthetic dataset resembling the ten classes of CIFAR-10 is created using latent diffusion, providing a direct comparison between real and AI-generated images. The classification task is framed as a binary problem, distinguishing real images from synthetic ones. To achieve this, a Convolutional Neural Network (CNN) is trained to classify the images with optimal performance. After fine-tuning hyperparameters and evaluating 36 distinct network architectures, the best-performing model achieves an accuracy of 92.98%. Additionally, explainable AI techniques, such as Gradient Class Activation Mapping, are applied to interpret the model’s decision-making process. Interestingly, the analysis reveals that rather than focusing on primary subjects, the model relies on subtle background inconsistencies to differentiate real images from synthetic ones. To support further research in this domain, the newly generated dataset, CIFAKE, is made publicly available for future studies.

DOI: /10.61463/ijset.vol.13.issue2.221