Image Resampling Detection via Spectral Correlation with False Alarm Control.

The detection of image resampling is a critical task in digital forensics, as it is essential for identifying manipulated media and verifying authenticity. Traditional approaches mostly rely on the correlations in the spatial domain caused by resampling, while the correlation in the Fourier domain related to the spectrum folding phenomenon during resampling is overlooked. In this work, we conduct a theoretical study of this phenomenon, and demonstrate that the spectral correlations caused by resampling are exploitable for image resampling detection. We propose an unsupervised method to detect anomalous spectral correlations, which consists of extracting image residuals to suppress natural correlations caused by the image content, and validating remaining significant correlations of resampling using an a contrario framework. The proposed method is suitable for images with a preserved aspect ratio, even if they are JPEG-compressed. Evaluated in diverse scenarios-including resampled images with different anti-aliasing filters, scaling factors and JPEG compression levels and in-the-wild resampled images from social media-the proposed method consistently outperforms existing methods, and shows robustness to different classical and AI-based interpolating filters and image sizes. Code is available at https://github.com/li-yanhao/ird.