AI Art Provenance Tracking for Tamper Detection

ISEF Category: Technology Enhances the Arts

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A fake image can look real even when every pixel has been edited. That makes provenance, the record of where an image came from, a big deal. If you can track every prompt, seed, model version, and edit, you can tell a cleaner story about what an artwork really is. Your project asks whether a local ledger can survive the kinds of edits people actually make.

What Is It?

This project is about proving where an AI-assisted artwork came from. Think of it like a receipt chain for an image. Each time the image is generated or changed, your system records the prompt, the model version, the random seed, and the edit history. A Merkle log is a data structure that links records together so one changed entry can break the chain. That makes silent tampering harder to hide.

The viewer-facing overlay is the part people can see. It could show a simple trust label, a timeline, or a warning if the file no longer matches the record. The key idea is not to judge whether the art is good. The key idea is to test whether the history is intact, even after normal image handling like cropping, compression, re-saving, or format changes.

Why This Is a Good Topic

This is a strong science fair topic because you can test it like an engineering problem. You have a clear output, tamper detected or not, and you can measure how well the system holds up under different edits. The project connects to real problems in art, media literacy, copyright, and misinformation. You can also learn about hashing, logging, metadata, and evaluation, which are useful skills in computer science and digital media.

Research Questions

• How does a Merkle-based provenance log compare with plain metadata for detecting edited AI-generated images?
• What is the effect of common image edits on tamper-detection accuracy?
• Does the type of edit, such as cropping, compression, or color adjustment, change how often the overlay flags a file?
• To what extent does stripping metadata reduce the system's ability to recover image history?
• Which combination of prompt, seed, and model version fields gives the most reliable audit trail for later verification?
• How does adding multiple edit checkpoints affect the chance of catching unauthorized changes?

Basic Materials

• Laptop or desktop computer with internet access.
• Python installed on the computer.
• Open-source image generation tool or access to a local generative model.
• Text editor or code editor such as VS Code.
• Free hashing library in Python.
• Spreadsheet software for logging results.
• Image viewer that can show file metadata.
• Folder system for saving original and edited images.

Advanced Materials

• University lab or departmental workstation with a GPU.
• Local open-source diffusion model or other image generation stack.
• Python with cryptography, image processing, and data analysis libraries.
• Database or structured file system for provenance records.
• Image forensics tools for comparing file fingerprints.
• Git or another version control system for protocol tracking.
• Dataset of artist-created and AI-generated images for stress testing.
• Statistical analysis software for tamper-detection benchmarks.

Software & Tools

• Python: Automates hashing, log checks, and analysis of tamper-detection results.
• ImageJ: Measures visible changes after cropping, compression, or color edits.
• VS Code: Helps you write and test the provenance script and overlay logic.
• Git: Tracks your code changes and supports reproducible development.
• Jupyter Notebook: Lets you explore results, make plots, and compare detection rates.

Experiment Steps

1. Define the exact provenance fields you will store for each image and decide which edits count as tampering.
2. Design the log structure, then choose how each record will be chained so one change breaks verification.
3. Build a test set of original images and planned altered versions that reflect common user edits.
4. Plan your verification rules so you can score each file as verified, altered, or uncertain.
5. Compare the system against simpler metadata-only tracking to see whether the ledger adds real protection.
6. Organize your evaluation so you can report detection rates by edit type, file format, and generation method.

Common Pitfalls

• Treating metadata as proof, when image files can keep or lose tags during re-saving.
• Testing only one edit type, which hides weak spots in the provenance system.
• Using a log that changes in place, which makes later tampering hard to detect.
• Forgetting to separate visual similarity from provenance verification, which makes the results hard to interpret.
• Skipping a baseline system, which leaves you unable to show that the Merkle log adds value.

What Makes This Competitive

A competitive version goes beyond a demo and becomes a real evaluation study. You would compare several edit types, several file formats, and at least one simpler baseline. Strong projects also report false positives and false negatives, not just a single accuracy number. If you test how the system behaves after common real-world handling, like export, compression, and metadata stripping, your project looks much stronger.

Project Variations

• Test the provenance system on AI-generated portraits, landscape art, or meme-style images to see which edits are easiest to trace.
• Compare a local append-only Merkle log with blockchain-style timestamping for image history verification.
• Add a viewer overlay that changes based on confidence level, then measure whether people understand the status faster.

Learn More

• NIST Digital Identity Guidelines: Search NIST for background on hashing, verification, and audit trails used in secure systems.
• MIT OpenCourseWare: Search for introductory computer security and cryptography lectures that explain hashes and integrity checks.
• PubMed: Search for review articles on deepfakes, media provenance, and digital forensics in scientific communication.
• ACM Digital Library: Search for peer-reviewed papers on image provenance, authenticity, and tamper detection.
• GitHub Docs: Read the free documentation on version control concepts if you want to track code and project history.