Deepfake Image Detection using Deep Learning

The Deepfake Image Detection system is a forensic-AI final-year project that answers one of the defining questions of the AI era: is this face real? Built with a convolutional neural network trained on the FaceForensics++ benchmark dataset, the system detects AI-manipulated faces — face swaps, reenactments and GAN-generated portraits — by combining spatial deep-learning features with frequency-domain artifacts that generative models leave behind.
The accompanying FakeSpot web app accepts an image upload, runs face detection and alignment, and returns a real/fake verdict with a confidence score plus Grad-CAM heatmaps that highlight the exact facial regions (eyes, mouth edges, blending boundaries) the model found suspicious — so the verdict is explainable, not a black box.
This is a built-to-order software project: the code, trained model weights and full documentation are prepared fresh for you, with the complete viva kit (report, PPT, Q&A) included.

1. You upload a photo (or video frames) to the FakeSpot web app.
2. The face detector locates and aligns every face in the image.
3. Each face crop is fed to the trained CNN (Xception/EfficientNet-style backbone), which scores spatial deep-learning features.
4. In parallel, the system computes the image's frequency spectrum (FFT/DCT) — generative models leave distinctive upsampling and blending artifacts that this branch flags.
5. The fused score produces a real/fake verdict with confidence; Grad-CAM back-projects the decision onto the face so you can show the examiner exactly which regions triggered the detection.

• Real/fake binary classification of faces with calibrated confidence score
• Dual-stream detection: spatial CNN features + frequency-domain (DCT/FFT) artifact analysis
• Grad-CAM explainability heatmaps that highlight manipulated facial regions
• Automatic face detection, cropping and alignment (OpenCV / MTCNN pipeline)
• FakeSpot web app — drag-and-drop image analysis in the browser
• Batch mode: scan a folder of images and export a results CSV
• Trained on the FaceForensics++ dataset (1,000+ videos across 4 manipulation methods)
• Training notebooks with loss/accuracy curves and confusion matrix included
• Viva-ready evaluation report: accuracy, precision, recall, F1 and AUC

• Complete source code (training, evaluation and inference, well-commented)
• Trained model weights (.pth) — demo works without a GPU
• FaceForensics++ data preparation and training scripts
• Project report PDF (problem statement, literature survey, architecture, results)
• PPT presentation for final review
• Viva Q&A preparation document (deepfake types, CNN layers, frequency artifacts, Grad-CAM, metrics)
• Setup guide (environment, dependencies, how to run training vs. inference)