AI Powered Fake Image Detection with Grad-CAM Explainability

Generative AI tools like Midjourney, DALL-E, and Stable Diffusion now produce fake images that look identical to real photographs. This makes it nearly impossible for humans to distinguish authentic content from AI-generated fakes by eye alone. The spread of such synthetic media poses serious risks including misinformation, identity theft, and erosion of public trust in digital content. A reliable automated detection system is therefore urgently needed. This paper presents a fake image detection system built to address this problem. The system uses three deep learning models working together. FaceNet, based on InceptionResnetV1 pretrained on VGGFace2, detects fake faces. EfficientNet-B4 pretrained on ImageNet detects fake building images. EfficientNet-B3 pretrained on ImageNet detects fake nature images. The user can select a specific model manually or use Auto-Detect mode. In Auto-Detect mode, MTCNN first looks for a face. If a face is found, the image routes to FaceNet. If no face is found, edge detection checks for straight building lines. Images with many straight lines go to the Buildings model. All other images go to the Nature model. An Ensemble mode runs all three models simultaneously and averages their P(fake) scores. This produces a consensus verdict that reduces individual model bias. For explainability, Grad-CAM generates heatmaps that highlight the exact image regions responsible for a FAKE verdict. The system was trained on 330,335 images containing both real photographs and AI-generated pictures across human portraits, architectural structures, nature scenes, and anime-style illustrations. The system runs as a Flask web application on localhost port 5001. Test results show detection of AI-generated building images at 67.2% confidence with P=0.672, and nature images at 69.2% confidence with P=0.692. All test samples were correctly classified with no false positives or false negatives.