AIGC Detector — AI-Generated Image Detection with Provenance Checks

Overview

AIGC Detector is a web service that estimates whether an image was produced by a generative model, and it is upfront about what that estimate can and cannot tell you.

How it works

A FastAPI web app handles uploads, URL fetching, format conversion (HEIC, BMP, TIFF), authentication and rate limiting. It forwards images to a serverless GPU backend on Modal, which runs a Swinv2 classifier on the full image and again on each tile of a 4×4 grid. Hot tiles inside an otherwise cool image point to local edits. Images are processed in memory and never written to disk.

Next to the score, the app checks the file's provenance: C2PA manifests are verified offline against vendored trust lists, and IPTC, EXIF and PNG generator tags are inspected. Provenance is reported separately and never folded into the score. FFT energy, noise-residual and JPEG-grid diagnostics are shown for explainability only.

Engineering notes

The verdict bands come from a pilot study rather than a default 0.5 cut-off, because many real photos score above 0.5. URL analysis is SSRF-safe (pinned IPs, per-redirect checks), and the rate limiter only trusts forwarded headers from known proxies. A benchmark harness on held-out real, FLUX and SDXL images showed that a two-model log-odds ensemble raises AUC from 0.919 to 0.956. The findings are written up in an 8-page IEEE-format paper with verified references.

The same backend also serves the synthetic-media seam of the VeriDoc fraud-detection platform.