Statistical modeling and likelihood ratio testing for resampling detection in TIFF images

Abstract

Resampling, including resizing, rotating, skewing, is a common technique in digital image tampering, typically used in conjunction with manipulations such as cloning or splicing to create visually seamless forgeries. Despite its sophistication, the resampling process inevitably leaves two main artifacts: (i) periodicity of resampled pixels, and (ii) variance incoherence between original and interpolated pixels. We exploit these artifacts to distinguish between authentic and resampled TIFF images though a two-step detection process: (i) analyzing and modeling characteristics of both authentic and resampled TIFF images, then (ii) developing statistical detectors by quantifying statistical deviations in these models. Compared to the current state-of-the-art methods, our contributions are threefold. First, we examine the complete processing pipeline, from a RAW image to a resampled TIFF image, to construct appropriate statistical noise models for both authentic and resampled images. Second, we leverage the periodic artifact to extract residual noise data and exploit their variance incoherence to develop (generalized) likelihood ratio test-based detectors for the resampling detection. Third, we derive closed-form expressions for the power function of the proposed detectors and provide an analytical performance evaluation. Numerical experiments on six well-known image databases using diverse interpolation kernels (i.e., nearest neighbor, linear, cubic convolution and cubic spline) validate the mathematical formulation of our detection approach and empirically demonstrate its superior performance.