Platform Drift in Social-Media Image Forgery Detection and Localization: An Application-Level Analysis

<title>Abstract</title> <p>Online social networks apply platform-specific upload--download pipelines such as recompression, resizing, chroma modification, and metadata rewriting that can distort forensic traces and reduce the reliability of image manipulation detection and localization. This paper presents a platform-drift robustness evaluation for a forensic system derived from the published REFORGE framework and a standardized visualization panel consisting of input, heatmap, mask, and overlay outputs. All metrics are computed on 20 matched samples per dataset at a fixed operating point in order to isolate platform-induced drift. Using four social platforms (Facebook, WeChat, Weibo, and WhatsApp) and four benchmark datasets (CASIA, Columbia, DSO, and NIST16), we quantify decision stability and three drift metrics: tamper-probability drift ($\Dpf$), mask-area drift ($\DArea$, in percentage points), and connected-component drift ($\DRegions$). Across all evaluated dataset--platform settings, binary decisions remain stable, while confidence and mask geometry exhibit platform- and dataset-dependent drift with rare outliers. We report global trends, dataset-specific sensitivities, and key outliers, and we conclude with deployment recommendations that motivate OSN-consistency regularization and platform-conditioned normalization as practical robustness enhancements.</p>