Technology / WORLD

MBZUAI and Khalifa University test FaceGuardian against AI image edits

The face-image method reduced identity similarity after edits by the diffusion models tested in the study.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

Researchers at MBZUAI and Khalifa University developed FaceGuardian, a method intended to make it harder for diffusion models to preserve a person’s identity while editing face images. It adds facial details that remain visually similar to the original for people but can confuse machine recognition, and is designed to withstand image resizing and compression.

In a celebrity-dataset experiment described by MBZUAI on 2 October, edited images were evaluated with face-recognition software. FaceGuardian recorded an FR similarity score of 0.272, compared with 0.833 for unprotected images; after compression, its score was 0.361, while the compared FaceLock method scored 0.709. The results apply to the open-source models and dataset tested by the researchers and do not guarantee protection with closed models or in every real-world condition.

Data

Timeline

FaceGuardian experiment

Evaluation of the identity-protection experiment

  1. · Occurred

    Research findings published

    The study reports model results under the synthetic or controlled benchmark conditions stated in the source.

Methodology

The unit registry has no matching unit for FR similarity scores; figures remain in the story text with dataset and comparison context. Results are specific to the source-stated model, edit and benchmark-dataset conditions; no cross-model or cross-country comparison or aggregation is made.

Updated:

Structured coverage

Subjects and places

Subject proposals

Geographic references

Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.

  • United Arab Emirates · Place role: Coverage · Place type: Country or area · Location precision: Regional
    Evidence
    • Explicit preserved editorial country/region; no city or national centroid inferredmbzuai.ac.ae ↗
Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading