CSAFE AI Altered Image Forensics

A dataset and tracking tool for studying how AI image-editing models alter photographs, and how those alterations can be detected. Developed by the Center for Statistics and Applications in Forensic Evidence (CSAFE).

The tracking tool is a Flask web app (app.py + tracker.html) that records allows users to record information about AI-altered images and runs automated forensic analysis on them. The ditial images and app code are stored in a GitHub repository. Image metadata records are stored in a shared Supabase database.

Setup

1. Clone the Repository

   Clone the GitHub repository to your computer. The records created with the app are stored in a Supabase database and the images themselves are stored in the GitHub.

2. Create a Free Supabase Account

   Go to supabase.com and create a free account. Ask a team member to add you to the AI-altered Images project under Project Settings → Team.

3. Install Python Dependencies

   Open a terminal and change directories to the AI-altered-image-forensics folder.

   cd path/to/AI-altered-image-forensics

   Create a virtual environment and install the required Python packages:

   python3 -m venv .venv
   source .venv/bin/activate
   pip3 install -r requirements.txt

NOTE: Depending on how python is installed on your computer, you might need to call python instead of python3 to launch the app. You also might need to call pip instead of pip3. NOTE: On Windows, activate the virtual environment with .venv\Scripts\activate instead.

4. (Optional) Install ExifTool

   If you want to use the metadata and EXIF features in the tracker app, you will need to install exiftool. If you don't want to install this app, the tracker app should still work but will not populate the metadata and EXIF fields.

5. Configure Supabase Credentials in Python

   Copy .env.example to .env:

   cp .env.example .env

   Then fill in your Supabase credentials in .env. You can get these by logging in to supabase.com

   • Open the AI-altered Images project, and go to Project Settings → Data API. The API URL is on the Data API tab. Copy and paste the API URL into the SUPABASE_URL field in .env. Delete any additional characters after "supabase.co".
   • Go to Project Settings → API Keys for the publishable key. Copy and paste the publishable key in the SUPABASE_KEY field in .env.

   SUPABASE_URL=https://your-project-id.supabase.co
   SUPABASE_KEY=your-publishable-key
   

6. (Optional) Configure AI or Not API Credentials in Python

   To use the AI or Not detector (see "Run AI or Not" below), also add an AIORNOT_API_KEY. Sign up at aiornot.com, open the dashboard, and create an API key under the API / Developer section (API access requires a plan that includes it). Paste the key into .env:

   AIORNOT_API_KEY=your-aiornot-api-key
   

   This key is optional — the rest of the app works without it, but the Run AI or Not button will report that the key is not set.

7. Start the app

   If your terminal is already in the CSAFE-AI-altered-image-forensics folder, activate the virtual environment and launch the app:

   source .venv/bin/activate
   PORT=5001 python3 app.py

   You need to activate the virtual environment each time you open a new terminal. On Windows, use .venv\Scripts\activate instead.

   Open http://localhost:5001 in a browser.

   NOTE: Port 5000 is blocked by macOS AirPlay Receiver; use 5001 or any other available port.

   NOTE: Depending on how python is installed on your computer, you might need to call python instead of python3 to launch the app.

   To stop the server, press Ctrl+C in the terminal where it is running.

Using the Tracking Tool

Every new original image is assigned a unique study ID with the format csafe_<###>. The study ID connects modified and altered images to the original image. Records appear in the left sidebar, grouped by the study IDs*.

Unsaved changes: If you switch to a different record, create a new record, or close/refresh the tab while a form has unsaved changes, the app should warn you before proceeding.

There are three image types:

1. Original — an unmodified image file as it came from the camera
2. Modified — a copy of an original (or another modification) that has been cropped, rotated, recompressed, or otherwise transformed; scene content is unchanged, except where cropping removes part of the frame
3. Altered — an original or modified image whose scene content has been changed, either by an AI model or manually in software such as Photoshop

Upload an original image

The app renames the original image with a unique study ID. Then the user records details about the image:

1. Click + New original in the sidebar.
2. Click Browse… next to the Original filename field and select the image from anywhere on your computer.
   • The app saves a copy of the image in real images/01-original/ and automatically assigns a unique study ID (csafe-001, csafe-002, …).
   • A renamed copy is created immediately in real images/02-original-renamed/ (e.g. csafe-001.jpg).
   • If an original image with the same filename already exists in the database, a warning is shown with the matching study ID instead.
3. Optionally, enter comments in Notes.
4. Click Save record. The image should now be listed in the sidebar.
5. Click the Analyze button in the Analysis results section at the bottom of the record.
6. (Optional) complete the manual sections of the Analysis results.

(Optional) Upload a modified image

If a modified version (cropping, resizing, recompression, etc.) of the original image was created and used as the input for the altered image, upload the modified image:

1. Apply the desired transformation externally (e.g. export at 80% JPEG quality in Preview). The modified image can be saved anywhere on your computer under any filename.
2. Click + New modification in the sidebar.
3. Click Browse… next to Select input image and select the source image, the renamed original or renamed previously modified image from which the new modified image was created.
4. Select the transformation type from Modification type (Cropped / Resized / Recompressed / Rotated / Other). The app auto-suggests a modified filename (e.g. csafe-001-recomp.jpg); edit the Modified image filename field if needed. On the next step, the app with save a copy of the image with this filename.
5. Click Browse… next to the Modified image filename field and select the modified image that you created. The app saves a renamed copy of the image in real images/03-modified/.
6. Describe the transformation in Modification details (e.g. exported in Preview at 80% JPEG quality).
7. Optionally, add additional context in Notes.
8. Click Save record.
9. Click the Analyze button in the Analysis results section at the bottom of the record.
10. (Optional) complete the manual sections of the Analysis results.

Upload an AI altered image

Upload and record details about an AI-altered image:

1. Generate an altered image using an AI tool or software such as Photoshop. Download or save the image somewhere on your computer with the filename suggested by the AI model or software.
2. Click + New alteration in the sidebar.
3. Click Browse… next to Select input image and select the source image, the renamed original or renamed modified image from which the new altered image was created.
4. Fill in the model details:
   • Model — select from the list of software and AI models. The list is populated automatically from the altered images/ subdirectories.
   • Version / variant — if you can find information about the version or setting chosen, record it (e.g. Grok-2, Gemini 2.0 Flash, Grok: I chose Quality instead of Speed)
5. Click Browse… next to the Filename as assigned by AI model field and select the downloaded or saved altered image. The app copies the image into altered images/<model>/downloaded/ automatically.
6. The Your assigned filename field auto-suggests a new name for the altered image appending -b<###> to the end of the source image's filename. The 3-digit number after the letter b is assigned sequentially per source image (e.g. csafe-002-b001.png, csafe-002-b002.png, csafe-001-recomp-b001.png).
7. Click Copy and Rename to create save a renamed copy of the altered image in altered images/<model>/renamed/.
8. Set the Date / time generated to when the image was produced.
9. Enter the exact text you gave the model in Prompt text and select the Prompt strategy (Vague / Specific / Adversarial).
10. Describe what was added or changed in Object added (e.g. hammer, handgun, knife).
11. Click the area(s) of the image that were modified in the interactive Region altered 3×3 grid.
12. Set Mask / selection used to Yes or No.
13. Rate the realism of the alteration in Subjective quality (1 = poor blend, 5 = convincing).
14. Record any observations about the result or the reason for your subjective quality rating in Notes.
15. Click Save record.
16. Click the Analyze button in the Analysis results section at the bottom of the record.

Browse the gallery

Click Gallery in the top bar to open a visual grid of all images, organized by study.

• Single-click a thumbnail to view the full image in a lightbox.
• Hover over a thumbnail to reveal the ⋯ button. Click it to open an action menu with options that depend on the record type:

Record type	Options
Original	View this record · Add new modification · Add new alteration
Modification	View original record · Add new alteration
Alteration	View original record

View this record / View original record closes the gallery and selects the record in the sidebar. Add new modification / Add new alteration closes the gallery, creates a new record of that type, and pre-selects the source image as the input.

Analyze an image

Each record has an Analysis results section at the bottom. Click Analyze to run the full forensic pipeline on the image associated with that record. Results appear immediately and are saved with the record.

Check	Method	What triggers a flag
EXIF / metadata anomalies	exiftool	Known AI software strings, missing camera fields, absent GPS
C2PA / Content Credentials	JUMBF block detection via exiftool	Embedded C2PA manifest; shows claim generator, actions, validation status
Metadata diff vs. input	exiftool on both images	Added / removed / changed fields (only when an input image is known)
ELA anomaly	Re-compress at quality=90, pixel diff	Max pixel difference > 15
Noise inconsistency	Block-level noise variance (numpy)	Uneven noise across 64×64 blocks
Compression blocking	DCT boundary analysis	Visible 8×8 block boundaries (JPEG only)

Record responses from AI detectors

The Analysis results section also collects responses from third-party AI detectors:

• AI or Not — click Run AI or Not to send the image to the AI or Not API (requires AIORNOT_API_KEY in .env). The decision (Likely AI / Likely real), the probabilities for human / AI / deepfake, and a per-generator class breakdown (e.g. Flux, GPT-4o, Stable Diffusion — which varies per image) are filled in automatically and saved with the record.
• Claude — paste results from Claude manually. Record the model / version, then pick a question from the Prompt dropdown (a fixed list of 10 standard questions, e.g. "Has this image been altered with AI?") and paste Claude's response. Each image stores a separate response per question, so you can switch questions in the dropdown and fill in answers one at a time.

Dashboard

Click Dashboard in the top bar to open aggregate analytics across all records, organized into collapsible groups:

• Summary — record counts, alterations by model, subjective-quality distribution, and a quality-by-model scatter plot.

• AI Indicators — indicator presence by model, visible-watermark coverage, and metadata tags by image type.

• AI or Not Detector — analysis of the AI or Not results recorded on each image:

  • Confusion matrix — treats "Likely AI" as the positive prediction and altered images as the positive ground-truth case, with originals and modified images grouped as the negative case. Cells show true/false positives and negatives, with the false positive rate and false negative rate below.
  • Detection rate by model — for altered images, the share flagged "Likely AI" per generating model.
  • AI probability distribution — density curves of the AI confidence score, overlaid for real (originals + modified) vs. altered images.

  Only records that have AI or Not results contribute; confusion-matrix cells and bars are clickable to open the matching images in the gallery.

• Visual / pixel-level artifacts — distributions of ELA and noise features.

• Random Forest Classifier — train and evaluate a classifier on the pixel features.

File naming conventions

Pattern	Meaning
csafe-001.jpg	Renamed original (study ID)
csafe-001-recomp.jpg	Modification — recompressed
csafe-001-cropped.jpg	Modification — cropped
csafe-001-recomp-b001.png	AI alteration — first altered version of the recompressed original
csafe-001-recomp-b002.png	AI alteration — second altered version

The b suffix stands for "bogus" (i.e. altered). The tracker auto-suggests the next sequential number when you create a new alteration record.

Adding a new AI model

The model dropdown in the alteration form is populated automatically from the subdirectories of altered images/. To add a new model:

1. Create a folder for the model inside altered images/, using a short lowercase name (e.g. comfyui):

   mkdir -p "altered images/comfyui/downloaded"
   mkdir -p "altered images/comfyui/renamed"

2. Commit the new folders to the repository so the model is available to all team members:

   git add "altered images/comfyui"
   git commit -m "Add comfyui model folder"
   git push

3. Restart the app (or reload the page if it is already running) — the new model will appear in the dropdown automatically.

The downloaded/ subfolder holds files as downloaded from the AI tool; the renamed/ subfolder holds the renamed copies created by Copy and Rename in the tracker.

Security

This is a local, single-user research tool with no built-in authentication. A few precautions keep the shared credentials and external connections safe.

• Enable Supabase Row Level Security (RLS). The SUPABASE_KEY is the publishable anon key, which is shared across the team. It is only safe if RLS is enabled on the records table with appropriate access policies. To set this up (a project owner/admin does this once):

  1. In the Supabase dashboard, open SQL Editor → New query.
  2. Paste the contents of supabase_rls.sql and click Run. This enables RLS and grants the anon role the four operations the app needs (select / insert / update / delete). The script is idempotent, so it is safe to re-run.
  3. Verify: the Database → Tables → records view no longer shows the red "Unrestricted" badge, and four policies appear under Authentication → Policies.
  4. Confirm the app can still load, save, and delete records.

  Never put the service_role key in .env — it bypasses RLS entirely. Note that RLS is a baseline (least privilege + no other tables exposed); it does not protect data from someone who already holds the shared anon key. True per-user protection requires per-user logins (Supabase Auth), a possible future enhancement.

• Protect your .env file. It holds your credentials and is already gitignored — never commit it. On a shared computer, restrict it to your user:

  chmod 600 .env

  If a key is ever exposed, create an new one (Supabase: Project Settings → API Keys; AI or Not: regenerate in your account dashboard).

• Run locally only. Start the app on the default localhost binding and do not expose the port to a network or bind to 0.0.0.0. The app has no authentication, and the Run AI or Not button spends real API credits on every call.

• Keep the debugger off. The app runs with the Flask debugger disabled by default. Only enable it for local development by setting FLASK_DEBUG=1, and never with the app reachable from a network — the interactive debugger allows arbitrary code execution.

• AI or Not sends images to a third party. Clicking Run AI or Not uploads the image to api.aiornot.com. Keep this in mind for sensitive material, and review their data-retention policy.

• Audit dependencies periodically. Check the pinned packages for known vulnerabilities:

  pip3 install pip-audit
  pip-audit

Troubleshooting

"Could not load records" error on startup

Free-tier Supabase projects pause automatically after a period of inactivity. When the project is paused, the app cannot reach the database and shows this error.

To restart the project:

1. Go to supabase.com and sign in.
2. Open the AI-altered Images project.
3. If the project is paused, you will see a banner at the top of the dashboard. Click Restore project.
4. Wait a minute or two for the project to fully restart.
5. Reload the app in your browser — records should load normally.

NOTE: Only a project owner or admin can restore a paused project. If you cannot restore it yourself, ask the team member who owns the project.

Repository structure

.
├── app.py                          # Flask backend
├── tracker.html                    # Single-page frontend
├── .env                            # Supabase credentials (not committed — get from a team member)
├── .env.example                    # Credential template
├── requirements.txt                # Python dependencies
├── supabase_rls.sql                # Row Level Security policies for the records table
├── migrate_to_supabase.py          # One-time migration script (already run)
├── real images/
│   ├── 01-original/                # Files as received from camera
│   ├── 02-original-renamed/        # Renamed copies (csafe-001.jpg, etc.)
│   └── 03-modified/                # Pre-processed versions (recompressed, cropped, etc.)
├── altered images/
│   ├── chatgpt/
│   ├── comfyui/
│   ├── gemini/
│   ├── grok/
│   └── ...
│       ├── downloaded/             # Files as downloaded from the AI tool
│       └── renamed/                # Renamed copies (csafe-001-recomp-b001.png, etc.)