"Semantic Manipulation Localization"

"The Semantic Manipulation Localization (SML) Dataset is a fine-grained benchmark designed for localizing meaning-altering semantic edits in multimedia images. Unlike conventional manipulation datasets that mainly focus on visible artifacts, boundary inconsistencies, or low-level forensic traces, this dataset targets subtle manipulations that preserve strong visual consistency while changing image interpretation. The dataset is constructed through a semantics-driven pipeline that identifies semantically decisive local regions, performs targeted semantic editing, and provides pixel-level localization masks. It also includes selected and synthetic cognition-relevant samples derived from public manipulation and saliency datasets, covering representative manipulation types such as splicing, copy-move, and removal. The dataset contains 23,632 image-mask pairs, including authentic images, manipulated images, and corresponding masks, supporting research on semantic-aware manipulation localization, multimedia content integrity, and trustworthy multimedia content understanding."