Joint statistics matching for camera model identification of recompressed images

Area of Science:

• Digital Image Forensics
• Computer Vision
• Signal Processing

Background:

• Source camera identification is crucial for digital forensics.
• Existing methods perform poorly on recompressed images due to significant accuracy degradation.
• JPEG recompression with varying quantization tables introduces substantial shifts in image features.

Purpose of the Study:

• To develop a robust method for source camera identification of recompressed images.
• To address the challenge of reduced accuracy caused by image recompression.
• To propose an algorithm that mitigates the impact of double JPEG compression.

Main Methods:

• Proposed a novel iterative algorithm: joint first and second order statistics matching (JSM).
• JSM learns a feature projection into a low-dimensional subspace.
• Simultaneously matches first and second-order statistics between training and test features, inspired by transfer learning.

Main Results:

• JSM effectively reduces feature divergence caused by recompression.
• The method preserves discriminative properties of the features.
• Extensive experiments on the Dresden Image Database demonstrated significant performance improvement over state-of-the-art methods.

Conclusions:

• JSM offers a principled dimensionality reduction approach for robust camera model identification.
• The algorithm successfully handles the challenge of recompressed images in forensic analysis.
• JSM significantly outperforms existing methods for identifying the source camera of recompressed images.