User guide for NIST Media Forensic Challenge (MFC) datasets
Open accessAbstract
More than 300 individuals from 150 organizations across 26 countries and regions use the NIST released Media Forensic Challenge (MFC) datasets for their research. The MFC datasets were created for use in the DARPA MediFor (Media Forensics) program. Since their release, multiple questions have been fielded regarding the dataset properties, including contents, metadata definitions, usage, data repurposing, etc. For example: what do the datasets contain? What are the definitions of the different kinds of metadata? How does one label the data with the reference information to build the training data for machine learning algorithms? How would one modify/extract the data for their own research purposes? This document serves as a user guide for the MFC datasets, including those used in the Nimble Challenge (NC). This guide includes: 1) a description about MFC datasets including background, evolution history, and the dataset summary by the evaluation tasks; 2) user access and permissions of MFC datasets; 3) an introduction to the MFC data by providing a simple example of a manipulation journal graph and its detailed corresponding MFC dataset reference files; 4) an introduction to a flexible subset selection approach, Selective Scoring, to sample the test probes from the entire test set for the particular task evaluation; 5) information to help users gain a deeper understanding of the metadata by presenting two commonly used approaches to illustrate the manipulation operation statistic histogram distributions, and 6) a general template of the NIST MFC evaluation dataset to facilitate the future dataset generation.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
OpenMFC describes detection tasks scored via receiver operating characteristic, area under the curve, and correct detection at a specified false alarm rate, such as CD@FAR = 0.05, aligning evaluation with thresholded decision-making.
“OpenMFC describes detection tasks scored via receiver operating characteristic (ROC), area under the curve (AUC), and correct detection (CD) at a specified false alarm rate (FAR), such as CD@FAR = 0.05, aligning evaluation with thresholded decision-making [5].”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [5]
Section 5.2 describes detection-system performance using ROC curves, including selective scoring for crop detection. The passages do not specify AUC or correct detection at a fixed false-alarm rate, including CD@FAR = 0.05, so the claimed threshold-specific scoring is not supported by this copy. (LLM, conf=0.86)
Experimental — accuracy not yet measuredOpenMFC frames recurring evaluation as necessary due to evolving manipulation technologies.
“It further frames recurring evaluation as necessary due to evolving manipulation technologies [5].”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [5]
The Introduction describes rapid advances in media falsification, and §2.3 describes successive evaluation datasets incorporating newer manipulation types, including GAN-based techniques. These passages support recurring evaluation in the context of evolving technologies, but the guide does not explicitly say that those advances make recurring evaluation necessary. (LLM, conf=0.82)
Experimental — accuracy not yet measuredOpenMFC explicitly argues that neutral benchmark evaluations are needed because system performance is sensitive to data/task variation and self-evaluation may not reflect field performance.
“OpenMFC explicitly argues that neutral benchmark evaluations are needed because system performance is sensitive to data/task variation and self-evaluation may not reflect field performance.”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [5]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 3 · Cited by 8
Research works printed but not linked: 1
- Fiscus, J. and Guan, H. (2020), Media Forensics Challenge Evaluation Overview, ARO Sponsored Workshop on Assured Autonomy, Workshop Talk. Available at https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=930628.
- Authors
- Guan, H., Delgado, A., Lee, Y., Yates, A. N., Zhou, D., Kheyrkhah, T., & Fiscus, J.
- Published
- 2021-07-06
- OpenAlex
- W3179810155
- Cited by
- 8
- References
- 3
- Citations by year
- 2026: 4 · 2024: 1 · 2023: 1 · 2021: 2
- Type
- report
- Language
- en
- Open access
- gold
- doi.org
- Funders
- National Institute of Standards and Technology
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
- Physical Sciences › Mathematics › Statistics and Probability › Benford’s Law and Fraud Detection
- Keywords
- multimedia forensics, metadata definition, training data labeling
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
