AntiFake · Real-time forgery detection engine

Tools that detect forgery after the fact already exist, but a late verdict is of little use. We built detection that completes while the verification process is still under way. This was a national R&D project under the Ministry of SMEs and Startups' Startup Growth Technology Development Program. It produced AntiFake, for which a patent has been filed.

Project delivery

Before

Started 2023.05

After

Completed 2024.07 · all 8 performance targets met

Background

Remote account opening, online identity verification, video interviews. As more processes identify people by their face, there is a growing need to confirm that the face is real. We addressed this problem as a project under the Ministry of SMEs and Startups’ Startup Growth Technology Development Program.

The problem

Technology for analyzing forgery after the fact already exists. This project targeted real time. If the verdict is not ready while the identity check is in progress, there is nothing to act on.

What we built

Technology that detects manipulated face video in real time
AntiFake, a software product released as a free app on the App Store, in 4 languages
One patent application filed in Korea

AntiFake machine learning screen. Tabs at the top read ‘Fake 탐지 / 머신러닝 / Update’ (Fake detection / Machine learning / Update). The ‘딥러닝 콘트롤’ (Deep learning control) panel on the left has a button to start training, and the ‘Deep Learning Status’ panel on the right shows a progress bar and training logs by epoch and iteration.

Beyond detection, the app also runs training on the device. New forgery techniques keep appearing, so training needs to continue in the field rather than waiting for a rebuilt model each time. The screen shows a training run in progress.

AntiFake in multiple languages. Three rows show English, Traditional Chinese, and Japanese from top to bottom, each with password recovery, sign-up, and login screens. Field names and button labels appear in each language. The start screen at the far right is blurred because of a person’s photo in the background.

The interface is available in 4 languages, including Korean. The image shows three of them, with field names and button labels in each language.

Results

The project was completed on schedule in July 2024. A patent for the result was filed in February 2024.

Performance figures were measured by an external testing body, Gumi Electronics & Information Technology Research Institute. It tested 8 items, and all 8 exceeded their targets. For example, mobile deepfake detection reached 59.2 FPS against a target of 7 FPS, measured on a Galaxy S21 Ultra. Test devices vary by item, so all 8 items and their test conditions are listed on the Technology page.

Related work is listed under AI adoption.