Reproduce “Noiseprint: a CNN-based camera model fingerprint” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Davide Cozzolino, Luisa Verdoliva

Researchers at Universit`a Federico II di Napoli developed Noiseprint, a CNN-based camera model fingerprint that extracts a highly discriminative noise residual by suppressing scene content and enhancing camera-specific artifacts. This method achieved state-of-the-art performance in blind image forgery localization, with an average MCC of 0.403 across diverse datasets, and showed promise for camera model identification.

Key insights

  • A neural network can learn to extract a robust "camera model fingerprint" by simultaneously suppressing scene content and enhancing subtle, model-specific processing artifacts, even without explicit ground-truth fingerprints.
  • The Siamese training paradigm, using positive and negative patch pairs with shared camera models and spatial positions, effectively teaches the network to distinguish model-specific characteristics implicitly.
  • The learned noiseprint, when fed into an existing blind localization algorithm, dramatically improves detection accuracy, demonstrating its superior quality and discriminative power over prior residual types.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.