AntiFake
Detects forged and altered face videos in real time. Developed under the Ministry of SMEs and Startups (중소벤처기업부) Startup Growth Technology Development Project (창업성장기술개발사업), with eight items measured by an external testing institution.
중소벤처기업부 창업성장기술개발(R&D) · Duration 2023.05 – 2024.07
A technology that detects forged and altered face videos in real time with AI,
developed to completion as a national R&D project.
Non-face-to-face account opening, remote identity verification, video interviews. As more procedures identify a person by their face, procedures that confirm the face itself is real have become just as necessary.
Technology that analyzes forgery and alteration after the fact already exists. What this project targets is real time. If the verdict does not come at the very moment the verification procedure is in progress, there is nothing left to do with it.
Numbers measured by an external institution.
The Gumi Electronics and Information Technology Research Institute (구미전자정보기술원, GERI) measured eight items, and all eight exceeded their targets.
| Evaluation item | Unit | Target | Measured | Verdict |
|---|---|---|---|---|
| Desktop measurements GPU RTX 3090 · CPU Intel Core i9-10900F · RAM 32GB | ||||
| Face image feature point extraction accuracy | % | 95 or higher | 99 | PASS |
| AntiDeepFake face extraction accuracy | % | 93 or higher | 99.84 | PASS |
| Face image feature point extraction time | ms | 33.3 or lower | 2.043 | PASS |
| AntiDeepFake face extraction time | ms | 222 or lower | 7.283 | PASS |
| Smartphone measurements Samsung Galaxy S21 Ultra (Android) · iPhone 13 Pro Max (iOS) · 5G | ||||
| Face image feature point extraction speed on mobile | FPS | 20 or higher | 70.3 Android 74.3 iOS | PASS |
| AntiDeepFake face extraction speed on mobile | FPS | 7 or higher | 59.2 Android 41.8 iOS | PASS |
| Processor usage during feature point extraction on mobile | % | 60 or lower | 46.8 Android 40.8 iOS | PASS |
| Processor usage during AntiDeepFake detection on mobile | % | 90 or lower | 48.2 Android 66.68 iOS | PASS |
- Testing institution
- 구미전자정보기술원(GERI)
- Test period
- 2024.03.29 – 2024.04.18
- Issue date
- 2024.04.26
- Test location
- 주식회사 브이로프
- Dataset
- FaceForensics++ (DeepFakes · Face2Face · FaceSwap · NeuralTextures)
- Measurement method
- Average of 30 samples × 5 repeated measurements
Where it applies.
-
eKYC · Non-face-to-face identity verification
Determines whether the face video submitted during account opening and identity verification is from an actual recording.
-
Remote identity verification
Checks whether the other party is being captured in real time during an identity verification conducted over video.
-
Video evidence verification
Applied to procedures that must determine whether submitted video material has been forged or altered.
-
Content platform filtering
Applied to services that must filter out forged or altered videos at the upload stage.
Available for free on the App Store.
| Released | 2024.02.06 |
|---|---|
| Supported languages | Korean · English · Traditional Chinese · Japanese |
| Availability | Downloaded in 25 countries (App Store, as of 2024.09.25) |
| Download | App Store |
Embedded use in corporate systems, technology transfer, and licensing are negotiated separately. Tell us about the environment you want to apply it to, and we will first lay out the feasible options.
Carried out as a national R&D project.
| Project title | AI 적용, 얼굴 위변조 영상 실시간 탐지기술 개발 |
|---|---|
| Support program | 중소벤처기업부 창업성장기술개발(R&D) |
| Specialized agency | 중소기업기술정보진흥원(TIPA) |
| Project period | 2023.05 – 2024.07 |
| Principal investigator | (CEO) |
| Total R&D cost | KRW 150 million (Government support KRW 120 million) |
| Development deliverables | AntiFake (Software) |
This technology is the result of the Startup Growth Technology Development Project (창업성장기술개발사업, Didimdol track) conducted by the Ministry of SMEs and Startups (중소벤처기업부).
We state the rights holder and the inventor together.
The CEO is listed as an inventor on all three.
| Type | Title · number | Filed | Registered | Rights holder · inventor |
|---|---|---|---|---|
| Registered | System and Method for Producing Special Effects through Face Recognition 10-2529209 | 2021.09.16 | 2023.04.28 | 주식회사 이엔터 Inventor and others |
| Registered | Distributed Processing System for Real-time Deepfake Image Detection in Mobile Environments 10-2896820 | 2022.12.28 | 2025.12.02 | 주식회사 이엔터 Inventor and others |
| Filed | AfterRemaker (Class 09) 10-2024-0022442 | 2024.02.16 | Under examination | 주식회사 브이로프 Inventor and others |
Collectors are built on the assumption that they will stop.
A collection run that spans days is bound to break somewhere in the middle. Responses stop coming, formats break, computers shut down. The stopping itself cannot be prevented. Instead, from the start we build in a mechanism that resumes from where it stopped and fills in only what was missed — so a run that stopped midway is never handed over half-done; we fill the gaps and deliver the rest.
-
01
We read the values the page already carries, not the screen
When you point at the screen structure with selectors, a site redesign can quietly bring in different values.
-
02
It remembers where it stopped
Large-scale collection is bound to break midway. When it restarts, it resumes from the saved progress point.
-
03
Re-fetching never overwrites what was already collected
It merges by unique identifier. Data already collected stays intact, and only duplicates are filtered out.
-
04
Failed items are kept in a list, and only those are re-fetched
Addresses that failed collection are recorded separately, and a script re-fetches only that list.
-
05
When it gets stuck, it switches paths and continues
When no response comes or a response arrives in a broken format, it switches to another path and continues from that point.
-
06
Handed over in a form that runs without a development environment
It is built as an executable so the person in charge can use it without installing Python.
The longest-running collection system we operated covered 40 or more sites for 5 years. It was not built, handed over, and done — this runtime comes from continually fixing it as target sites changed or broke.
Technologies worked with across 23 years of development
The CEO directly designs and implements. Depending on project scale, developers are directly selected to join, and requirements are never subcontracted wholesale to third parties.
- Python
- PHP
- Linux
- HTML
- Docker
- AWS
- EC2
- GitHub
- Redis
- MongoDB
- Node.js
A permanent team of 3 works directly, with no subcontracting to third parties.
If you are considering adoption or technical collaboration.
Tell us the target environment and what needs to be detected, and we will first lay out which approaches are feasible.