Synthetic Image Detection using Deep Learning

The proliferation of AI-generated synthetic images produced by Generative Adversarial Networks (GANs) and diffusion-based models has created critical challenges for digital forensics, media authentication, and public trust. This paper presents a complete Synthetic Image Detection (SID) system capable of classifying any digital image as REAL or AI-generated without requiring access to any external dataset at inference time. The system combines two complementary deep learning architectures — a fine-tuned ResNet50 Convolutional Neural Network and a Vision Transformer (ViT-B/16) — both trained on the CIFAKE benchmark comprising 120,000 images. ResNet50 achieved 98.08% test accuracy with AUC-ROC of 0.9971, and ViT-B/16 achieved 98.05% accuracy with AUC-ROC of 0.9984, both substantially exceeding the published CIFAKE baseline of 92.98% (Bird and Lotfi, 2024). GradCAM++ (Gradient-weighted Class Activation Mapping++) was implemented from scratch to generate pixel-level visual explanations identifying the specific image regions influencing each classification decision. Beyond deep learning, a novel multi-signal forensic analysis pipeline was developed incorporating six independent pixel-level detectors: Discrete Fourier Transform checkerboard analysis, Error Level Analysis, SRM noise residual inspection, colour channel statistical analysis, Local Binary Pattern texture entropy, and Haar wavelet energy ratio computation. All six signals are combined with the two deep learning models in a weighted mega-ensemble producing a final verdict purely from the uploaded image pixels. The complete system is deployed as a RESTful API using FastAPI with a five-page Streamlit web interface providing real-time detection, GradCAM++ visualisation, and forensic signal breakdown.