Back to projects

Research

AI-Image Forensics

Which simple features expose AI-generated images?

PythonComputer VisionForensicsJPEGPre-registration
Repo

Role

Independent Researcher

Duration

2026

Date

2026

Type

Independent Research

By the Numbers

5

Pre-registered Experiments

Locked v1.0

Pipeline Spec

In-repo

Final Manuscript

Overview

A measurement study, explicitly not a new detector: five pre-registered experiments (E1-E5) mapped to four hypotheses test which explainable feature set (pixel-pattern, frequency-spectrum, or combined) best separates real photographs from AI-generated images. The pipeline specification was locked at v1.0 (stages 0-10) before results ran, with a deliberate negative-control experiment for mismatch shortcuts, JPEG quality levels 90/75/50, and an unseen-generator holdout, culminating in the headline E5 condition (unseen generator + JPEG 75). The repo carries the final results, a separate limitations report, a 20-paper curated reference set with BibTeX, and a CITATION.cff, alongside a written scope statement of what the study is not.

Tech Stack

PythonComputer VisionForensicsJPEGPre-registration

Inside the Project

What it does, how it works, and what it ships.

lock

Locked Before Results

The experimental pipeline was frozen at v1.0 with a self-checking figure script, so the analysis could not drift to fit the outcomes.

rule

Negative Controls

A deliberate mismatch-shortcut experiment (H2) checks that the classifier is not cheating through generator fingerprints, alongside the unseen-generator holdout.

  • JPEG 90 / 75 / 50 robustness ladder
  • Strict split by image group
star

Complete Research Layout

  • reports/final_results.md and reports/limitations.md
  • 20-paper curated reference set with BibTeX
  • CITATION.cff, LICENSE, per-experiment protocol/code/results/analysis

Outcomes

A finished measurement study with every artifact a reviewer would ask for.

image_search

E1-E5

Pre-registered Experiments

article

Final

Manuscript In-Repo

menu_book

20

Curated References

Interested in this project?

Reach out for source code walkthroughs or collaboration.

RepoView All ProjectsContact