Abstract: Document source detection uses unique copies to identify sources of leaked documents. The unique copies are generated from an original document and include a unique watermarking of one or more perturbations to a feature of the original document. An artifact, such as a photo or copy, is derived from one of the unique copies, the unique copy from which it was derived is identified. To identify the unique copy, unique copy keypoints in the unique copies are matched to artifact keypoints in the artifact to align the artifact with a location within the unique documents. Pixel regions in the unique copies that include perturbations are used to identify corresponding pixel regions in the artifact. Pixels in these regions are compared to identify the unique copy from which the artifact was derived, thus identifying a possible source of the leaked document.

Abstract: Unique copies of an original document can be generated and provided to individual recipients. The unique copies can be used to identify the source of a document leak. The unique copies are generated by replacing terms within the original document with alternative terms. The alternative terms are determined using a first machine learning model that receives a term from the document and outputs the alternative terms. The output alternative terms are provided to a second machine learning model that indicates a tone for each alternative term. The tone of the alternative terms is compared to the tone of the term from the original document, and one or more of the alternative terms are selected based on the tone of the alternative terms relative to the tone of the document term. The alternative terms used to generate the unique copies have a same or similar tone as the document term.