Deepfake Shield
한국어ReportDetection
NewLocation v14 · Image v7 · Video v6

Faces can be borrowed.
Traces remain.

Upload a file and detection starts right away. One report shows whether it is AI-generated, whose face it shows and where it was manipulated.

Detection See how it works →

Images 5~9 sec · Videos take longer depending on length · JPG PNG WEBP MP4 MOV · Up to 200MB · Use only public material or your own files in this demo

Deepfake Shield — Veritas odit moras

Measured on public test data

  • OpenSDIAI edit location
  • TGIFAI edit location
  • BAZEImage-to-video
  • KorusTraditional compositing

95.1%

Video detection accuracy · 1.8% false positives

Sealed test set of 206 videos

1.3%

Real photos wrongly flagged as AI · 92.4% of new-generator images detected

320 official Instagram photos

87.1%

Manipulated regions found · 68.5% on natural edits we made ourselves

Public edit dataset, OpenSDI 498 images

5~9sec

One image · Videos take longer depending on length

With the server warmed up

One detection engine, four uses

From a single image to a pool of videos, the same engine applies the same standard.

Image

Image detection

Shows whether it is AI-generated and where it was manipulated, with a score and a heatmap.

Learn more →
Video

Video detection

Picks a key frame from each scene and checks it with captions removed.

Learn more →
Dome

Enrolled people scan

Enroll people with consent, and the video pool is searched for them first.

Learn more →
Report

Reports and review history

Detection evidence sits on one page, along with the review history.

View report →

Detection takes three steps, and Dome also reads ad text. Each step leaves evidence in the report.

FIG 0.1

Is it AI-generated?

After on-screen text is removed, two different detection models score each scene together.

FIG 0.2

Whose face is it?

Faces are compared with people enrolled with consent, and the name and match score are recorded.

FIG 0.3

Where was it manipulated?

The frame is split into small patches to mark suspected manipulated regions and measure their area.

FIG 0.4

What is being advertised?

Dome scans read on-screen text and record what is being promoted.

How one video becomes a report

Dome finds faces first, then checks only the matches: AI or not → suspected manipulated regions → what is being advertised. Scroll down to build it up step by step.

A scene from a short investment ad filmed in a car. The person is fictional, made for this sample.

public-sample-02.mp4 · Video · 0:14Sample

Faces
Enrolled person A · Match score 0.41—
AI score
0.841 · Suspected AI—
Suspected regions
Face area · 11.4% of the frame—
Ad text
Investment advice · Guaranteed high returns—
Report
Pending review—

  1. 01 / 05 · Face matching

    Enrolled faces come first

    Faces are first compared with people enrolled with their own or their agency's consent. High match scores are named automatically, and unclear ones go to human review.

  2. 02 / 05 · AI detection

    Captions removed,
    two models check together

    Two different detection models score each scene after on-screen text is removed. Clearly high scores are marked Suspected AI, and unclear ones go to a person.

    • Accuracy on 206 sealed test videos95.1%
    • False positives on 240 human-filmed Instagram videos0.4%
    • Commercial generator videos7 of 12
    • Google Flow videos7 of 7

  3. 03 / 05 · Suspected regions

    Where it was manipulated
    is marked on the frame

    The frame is split into small patches, each is scored for manipulation, and high-scoring areas are marked. Videos are marked only with clear signs of AI, and images are also marked when a partial edit is suspected. Face swaps may mark the face, and fully generated videos may mark the whole frame.

    • Manipulated regions found on a public edit dataset87.1%
    • Natural edits we made ourselves68.5%

  4. 04 / 05 · Ad text

    Reads what is being sold

    With ads that borrow a celebrity's face, what is being sold is key. Text in each scene's key frame is read, and the product and wording go into the report.

  5. 05 / 05 · Report

    A person reviews it,
    and it stays on record

    Scores, the manipulation heatmap and on-screen text sit in one report. Nothing is confirmed automatically, and the reviewer, date and notes stay on record.

    View report →

What Shield does, line by line

Learn from this person:
the KRW 10M-a-month secret

On-screen text removed → Only the remaining scene is checked

0:020:050:090:12

1 frame per scene

1 image → 6-sec video · Suspected AI

2 suspected regions · 6.8% of the image

Compared only with people enrolled with consent

Inconclusive · Goes to a person

Inconclusive band · Not confirmed automatically

  • cosmetics-ad.pngSamplePending review
  • public-sample-03.pngSamplePending review
  • Street video · 0:08Face 0.29Pending review

Removes captions firstScenes are checked with text removed so on-screen text doesn't sway the score.

Deepfake Dome · Enrolled people scan

Enroll the people to protect, and the video pool
is searched for them first

Enroll a photo, name and affiliation. The pool is searched for that face, and only matching videos move on to detection.

Open Dome →

The first scenes of eight short videos in the pool, with one marked as matching an enrolled face. This scene is a sample.

Only people who consented

Only people who consented to scanning, directly or through their agency, can be enrolled. Enrolled faces are stored only on this server.

Private individuals are not searched

Private individuals' faces are never searched across the internet. Only the video pool is searched.

Only matches move on

Only videos with a face match go on to AI detection, suspected regions and ad text, in that order.

When to use it. Each situation calls for a different first look.

Impersonation ads

Finds investment and cosmetics ads that borrow a celebrity's face, and records the product and wording in a report.

Agencies · Platforms

Reports of non-consensual synthetic videos

Shows whether a victim's video is synthetic and where it was manipulated, to support counseling and takedown requests.

Victim support groups

Pre-publication checks

Checks whether tipped images and videos are AI-generated before they are reported.

Newsrooms

Report triage

Ranks reported files by score so people see the most urgent first.

Platforms

Protecting your talent

Searches the video pool first for videos showing people enrolled with consent.

Agencies · Management

Evidence for takedown requests

Attach a one-page report with the score, heatmap and on-screen text to takedown requests.

Support groups · Legal

Update history

Always improving, version by version

We find weak spots, fix them and record what changed in each version.

  1. Location v14 · Nov 7

    Finds more, paints less on real photos

    On 200 natural edits we made ourselves, fully found images rose from 84 to 105, and real photos painted fell from 2 to 0. Checking at two scales adds a few seconds per image.

  2. Image v7 · Face-swap location · Nov 5

    Catches new-generator images

    Detects 208 of 225 images made with gpt-image-2 (v3 caught 55). Swapped faces are re-checked on their own.

  3. Location model v6

    Finds more natural edits

    Finds 90.2% on a public edit dataset and 65.5% on natural edits we made ourselves.

  4. Video v6 · Image v3

    Fewer false positives first

    We chose versions that first reduce real content being flagged as AI: 0.4% false positives on 240 human-filmed Instagram videos and 1.2% on real photos.

  5. Face blur

    Faces in results are blurred automatically to protect privacy

    Faces are blurred in result images and in every video frame.

  6. Detection server

    Less waiting

    Models load when the server starts.

Sensitive files, kept in-house

  1. 01

    Files stay on the server

    Uploaded files and results stay on this server and are deleted after the demo.

  2. 02

    No outside services or chat AI

    Detection uses our own models only. Nothing goes to another company's servers.

  3. 03

    Only people who consented are searched

    Only people who consented, directly or through their agency, can be enrolled. Private individuals' faces are never searched across the internet.

  4. 04

    Nothing is confirmed automatically

    A report is evidence. Final decisions and legal judgments belong to the people in charge and their lawyers.

Frequently asked questions

Where do uploaded files stay?

Only on the demo server, and they are deleted after the demo. Detection doesn't use outside services or chat AI, so nothing is sent elsewhere.

Can it search for anyone's face?

No. Only people who consented to scanning, directly or through their agency, can be enrolled, and enrolled faces are stored only on this server. Private individuals' faces are never searched across the internet.

Does a high score prove it is fake?

No. High scores are marked "Suspected AI", and unclear cases are set aside as "Inconclusive". A report is supporting evidence, and final decisions and legal judgments belong to the people in charge and their lawyers.

What files can I upload?

JPG, PNG and WEBP images and MP4 and MOV videos, up to 200MB. Images take 5~9 sec, and videos take longer depending on length.

Can it miss things?

Yes. It can miss partly edited images and images from new generators. It caught 7 of 12 videos from commercial generators. "Not detected" does not mean it is real.

Does it find Photoshop composites?

The current location model is trained on regions manipulated by AI. It is still weak on traditional paste-and-erase composites, and we are improving that.

Check the traces first.

Detection View report →