MFC Datasets: Large-Scale Benchmark Datasets for Media Forensic Challenge Evaluation

摘要

We provide a benchmark for digital Media Forensics Challenge (MFC) evaluations. Our comprehensive data comprises over 176,000 high provenance (HP) images and 11,000 HP videos; more than 100,000 manipulated images and 4,000 manipulated videos; 35 million internet images and 300,000 video clips. We have designed and generated a series of development, evaluation, and challenge datasets, and used them to assess the progress and thoroughly analyze the performance of diverse systems on a variety of media forensics tasks in the past two years. In this paper, we first introduce the objectives, challenges, and approaches to building media forensics evaluation datasets. We then discuss our approaches to forensic dataset collection, annotation, and manipulation, and present the design and infrastructure to effectively and efficiently build the evaluation datasets to support various evaluation tasks. Given a specified query, we build an infrastructure that selects the customized evaluation subsets for the targeted analysis report. Finally, we demonstrate the evaluation results in the past evaluations.