Camera model identification based on the heteroscedastic noise model

Area of Science:

• Digital Image Forensics
• Statistical Signal Processing

Background:

• Accurate identification of camera models is crucial for digital image forensics.
• Existing methods often rely on simplified noise models that do not fully represent natural raw images.

Purpose of the Study:

• To design a robust statistical test for camera model identification.
• To leverage a heteroscedastic noise model for improved accuracy in identifying unique camera fingerprints.

Main Methods:

• The study frames the camera model identification problem within hypothesis testing theory.
• A likelihood ratio test (LRT) is developed for ideal scenarios with known parameters.
• Two generalized LRTs are proposed for practical applications with unknown parameters, ensuring controlled false alarm rates.

Main Results:

• Theoretical performance of the LRT is established under ideal conditions.
• Generalized LRTs demonstrate high detection performance while meeting false alarm probability requirements.
• Validation on simulated and real raw images confirms the approach's effectiveness.

Conclusions:

• The proposed statistical test based on the heteroscedastic noise model provides a reliable method for camera model identification.
• The two-parameter fingerprint derived from the noise model offers a unique identifier for different camera models.
• The generalized LRTs offer a practical solution for real-world image forensic applications.