An interactive web application that detects whether a video or image is real or AI-generated (deepfaked). Built with a custom CNN-GRU deep learning model (ResNet18 backbone + bidirectional GRU) trained to analyze temporal and spatial inconsistencies across video frames.
year
'25
CNN
ML
Digital Image Processing
Problem
The rise of accessible deepfake generation tools has made it increasingly difficult for everyday users to distinguish manipulated media from authentic content, fueling misinformation, identity fraud, and erosion of trust in visual media. Most existing detection tools are either locked behind research papers/notebooks, require technical setup, or aren't accessible to non-technical users who need a quick, intuitive way to check suspicious content.
Solution
I built an end-to-end deepfake detection system with a simple, general-purpose interface: users upload any image or short video (or try built-in real/fake examples) and receive an instant Real/Fake prediction with a confidence score. The model processes video by sampling frames evenly across the clip, extracting spatial features per frame with a CNN backbone, and modeling temporal patterns across frames with a GRU to catch inconsistencies that single-frame detectors miss. The app is fully deployed and publicly accessible on Hugging Face Spaces, so anyone can test it without installing anything — no local setup, no API keys, just open the link and try it.
