Exploring neural network-based approaches for removing device-specific camera fingerprints (sensor pattern noise) from images, addressing privacy risks in digital forensics.
Lead Researcher: Qile Zhang
Advisor: Prof. Zhongjie Ba, Zhejiang University
Period: May – June 2024
Every digital camera leaves a unique fingerprint — known as sensor pattern noise (SPN) — embedded in every image it captures. While useful for forensic source identification, this fingerprint poses a serious privacy risk: an adversary can link images to a specific device and, by extension, to its owner. This project explores whether neural networks can effectively suppress camera fingerprints while preserving image quality.
We designed and evaluated multiple strategies for fingerprint removal:
| Method | Description |
|---|---|
| Autoencoder (AE) | Unsupervised fingerprint suppression — the model learns to reconstruct images while implicitly removing device-specific noise patterns |
| Proxy-based Training | A training strategy to handle the absence of paired clean/fingerprinted data, using proxy fingerprints to supervise the removal process |
| Diffusion-based Methods | Evaluated diffusion models as an alternative denoising approach for fingerprint suppression |
| CNN Baselines | Compared against standard CNN-based denoising architectures |
- The autoencoder-based approach effectively suppresses camera fingerprints while maintaining acceptable image quality
- Proxy-based training successfully addresses the challenge of missing paired training data
- Trade-offs exist between fingerprint removal strength and image fidelity — more aggressive removal can introduce artifacts
├── Code/ # Implementation of all methods
├── Project——report.pdf # Detailed project report
└── README.md
cd Code
python main.py # See Code/ for detailed instructionsThis project was conducted at Zhejiang University under the supervision of Prof. Zhongjie Ba.