AI-powered image authenticity & deepfake detection — runs locally first, then secure cloud validation
Supports JPG, PNG, WEBP, HEIC · Max 20MB
Awaiting Source Image...
It's mixed, not fully local. EXIF metadata extraction (camera model, GPS coordinates, editing software) runs entirely in your browser — that part never leaves your device. The AI-based checks (deepfake detection, Error Level Analysis, manipulation detection) upload your image over an encrypted connection to AWS Textract and Rekognition, because that analysis genuinely requires server-side models. It's processed immediately and not stored beyond the session.
The tool performs several independent checks: EXIF metadata inspection (camera model, GPS coordinates, editing software), Error Level Analysis to highlight regions with inconsistent JPEG compression (a common indicator of compositing or local edits), and visual artifact inspection to spot copy-move manipulation.
Error Level Analysis re-compresses an image at a known JPEG quality level and computes the pixel-by-pixel difference between the original and re-compressed versions. Areas that have been digitally inserted or edited typically retain a different compression level than the unmodified background, appearing as brighter regions in the ELA visualisation.
ELA detects image manipulation by re-compressing a JPEG at a known quality level and comparing it to the original. Authentic images show uniform error levels across similar surface types. Edited regions typically show different compression artifacts and appear brighter in the ELA visualization, indicating they were saved at a different compression level — a sign of manipulation.
EXIF data is embedded in JPEG and TIFF files by cameras and phones. It contains: camera model, GPS coordinates, date/time, lens settings, and software used. Removed or inconsistent EXIF data (e.g., no GPS on an outdoor photo, or editing software tags) can indicate the image has been processed. Entirely missing EXIF is common in screenshots or web-downloaded images.
The AI model analyzes pixel-level patterns, lighting consistency, shadow geometry, face symmetry, and texture coherence. AI-generated images (deepfakes, GAN outputs, diffusion models) often show subtle artifacts in hair, teeth, backgrounds, and reflections that the model has been trained to detect.
No single indicator is definitive. A high ELA anomaly score with inconsistent EXIF and low AI authenticity together strongly suggest manipulation. But high ELA can also occur with legitimate heavy re-editing for artistic purposes. Treat the analysis as probabilistic evidence, not forensic proof. Always corroborate with other signals and context.
EXIF metadata extraction runs entirely in your browser and is never sent anywhere. For the AI-based checks (deepfake detection, ELA, manipulation detection), your image is uploaded over an encrypted connection to AWS Textract and Rekognition — those checks genuinely require server-side models and can't run client-side. Images are processed immediately and not stored beyond the session, and are never used for training or analytics.