The forgery verdict comes at the very moment identity verification is in progress | VROF

Background

Non-face-to-face account opening, remote identity verification, video interviews. As procedures that identify a person by their face multiply, so does the need for a procedure that verifies whether that face is real. We took this on as a 중소벤처기업부 (Ministry of SMEs and Startups) 창업성장기술개발사업 (Startup Growth Technology Development Program) project.

The problem we defined

Technology that analyzes whether something is forged after the fact already exists. What the project aimed for was real time. If the verdict does not come at the very moment the identity verification procedure is in progress, there is nothing you can do with the verdict.

What we built

Technology that detects face-forgery videos in real time
The software product AntiFake — released on the App Store as a free app, supporting 4 languages
1 domestic patent application

AntiFake app’s machine-learning screen. At the top are the ‘Fake 탐지 / 머신러닝 / Update’ tabs; on the left, ‘딥러닝 콘트롤’ (deep-learning control) has a start-training button, and on the right, ‘Deep Learning Status’ shows a progress bar with per-epoch and per-iteration training logs.

Beyond detection, training also runs directly inside the app. Instead of rebuilding and shipping a model every time a new forgery method appears, we believed training had to be able to continue in the field. What you see on the screen is that training in progress.

AntiFake app’s multilingual screen. It is divided into three rows — English, Traditional Chinese, and Japanese from top to bottom — with password recovery, sign-up, and login screens in each row. Field names and button labels appear in each language. The start screen at the far right is blurred because of the person photo laid out behind it.

The screens appear in 4 languages including Korean. The figure above shows three of them; field names and button labels change per language. The rightmost column is the start screen, which had a test person’s photo in the background, so it is blurred — for the same reason the cover screen was cropped to the verdict panel only.

Result

The project was completed as planned in July 2024. A patent was filed for the outcome in February 2024.

The performance figures come from an external testing institution. 구미전자정보기술원 (Gumi Electronics & Information Technology Research Institute) measured 8 items, and all 8 exceeded their targets. For example, mobile deepfake detection speed came out at 59.2FPS against a target of 7FPS, measured on a Galaxy S21 Ultra. Since the measuring equipment differs per item, all 8 items and their measurement conditions are listed as-is on the technology page.

Other work on the same subject is collected under AI 적용 (AI application).