Image Resampling Detection via Spectral Correlation with False Alarm Control

Area of Science:

• Digital Forensics
• Image Processing
• Signal Analysis

Background:

• Image resampling detection is crucial for verifying media authenticity.
• Existing methods often overlook Fourier domain correlations, focusing on spatial domain artifacts.
• Spectrum folding during resampling creates exploitable spectral correlations.

Purpose of the Study:

• To theoretically investigate Fourier domain spectral correlations caused by image resampling.
• To develop an unsupervised method for detecting image resampling using these spectral correlations.
• To demonstrate the effectiveness of the proposed method in diverse scenarios.

Main Methods:

• Theoretical analysis of spectrum folding phenomenon in the Fourier domain.
• Extraction of image residuals to isolate resampling-induced correlations.
• Application of an a contrario framework to validate significant spectral correlations.
• Unsupervised detection of anomalous spectral correlations.

Main Results:

• Spectral correlations caused by resampling are exploitable for detection.
• The proposed unsupervised method effectively detects resampled images.
• The method shows robustness against various anti-aliasing filters, scaling factors, JPEG compression, and interpolating filters (classical and AI-based).
• Outperforms existing methods in diverse evaluation scenarios, including social media images.

Conclusions:

• The developed method provides a robust and effective approach for image resampling detection.
• Exploiting Fourier domain correlations offers advantages over traditional spatial domain methods.
• The technique is suitable for images with preserved aspect ratio, including JPEG-compressed ones.