Face Analysis Technology Evaluation (FATE) MORPH
Latest Report | Participation Agreement | API | Validation | Encryption | Submit
Overview
This page summarizes and links to all FATE data and reports related to face morphing.2025-08-18 NISTIR 8584 FATE MORPH Part 4B: Considerations for Implementing Morph Detection in Operations PDF
2026-04-07 NISTIR 8292 Draft Supplement: FATE Part 4: MORPH - Performance of Automated Face Morph Detection PDF
2022-07-28 NISTIR 8430: FATE MORPH Part 4A: Utility of 1:N Face Recognition Algorithms for Morph Detection PDF
Status
A new demorphing track has been added to our evaluation, which tests algorithmic ability to recover images of original identities from a morphed photo. Updates have been published in the API document and validation package.
[2026-05-29] An updated FATE MORPH report has been published, adding a new plot of face recognition vulnerability at an elevated threshold of FMR=0.000001, demonstrating that for a number of face recognition algorithms, morphing vulnerability is notably lower at a more stringent threshold, with nominal increase in false non-match rate. Additionally, we’ve started reporting face recognition vulnerability at a threshold targeting FMR=0.000001 here.
[2026-04-07] An updated FATE MORPH report has been published, adding results for one new algorithm submitted by University of Twente (utw-000).
[2026-02-27] An updated FATE MORPH report has been published, adding results for two new diffusion-based morphing datasets. Results for these datasets are published for algorithms submitted in 2022 and later.
[2026-01-09] An updated FATE MORPH report has been published, adding results for one new algorithm submitted by secunet (secunet-004).
[2025-08-18] We have released NISTIR 8584 - FATE MORPH Part 4B: Considerations for Implementing Morph Detection in Operations. As described in this press release, the document is intended to build awareness and to guide organizations toward effective deployment of tools and practices in situations where morphed photographs are a concern in operational workflows. It includes guidelines for what organizations might consider doing after a morph detector generates a positive indication or a suspicious photo is detected through human review.
Accuracy Summary
[last updated: 2026-05-29]
Morph Detection Performance
Morph Detection Performance
The table provides a summary of all algorithms measured on morphing attack classification error rate (MACER) when bona fide sample classification error rate (BSCER) is set to 0.003, across a subset of the different morphing datasets used in our evaluation. MACER, or morph miss rate, is the proportion of morphs that are incorrectly classified as bona fides (nonmorphs). BSCER, or false detection rate, is the proportion of bona fides falsely classified as morphs. Timing durations were computed on an Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz machine for algorithms submitted after February 2022.
Face Recognition Accuracy on Morphs
Face Recognition Accuracy on Morphs
The graphs below plot face recognition algorithm vulnerability on morphs against general algorithm accuracy on non-morphed photos. Each circular point represents a face recognition algorithm recently submitted to the NIST FRTE 1:1 activity, and each triangular point represents a face recognition algorithm submitted to the NIST FATE MORPH activity. Note that algorithms submitted to FRTE 1:1 are not necessarily designed to handle morphed photos, and results are presented only as a point of reference. Submissions to FATE MORPH are designed with goals of face recognition algorithm resistance against morphing. The y-axis plots MMPMR, which is the fraction of morphs where both subjects incorrectly match to the morph. The x-axis plots FNMR or miss rate on regular photos, which provides an indication of general algorithm accuracy. Both MMPMR and FNMR are calculated with thresholds set to where the false match rate (FMR) is 0.0001 and 0.000001 respectively. The lower the MMPMR, the better the algorithm performs against morphs. MMPMR is notably lower at a more stringent threshold targeting FMR = 0.000001, and depending on the algorithm, with nominal increase in FNMR.
The table below provides numerical tabulation of MMPMR and FNMR for recent face recognition algorithms submitted to FATE MORPH and FRTE 1:1, ordered initially by FNMR.