evidoria

← Back to browse

Good practice Imported

South Korea's AI Deepfake Detection Model for the 2026 Local Elections

South Korea · Sejong · See the South Korea profile · See the Sejong profile

Evidence: Descriptive / self-reported Top 62% 53/100 · Ask Evidence Copilot about this practice

South Korea's Ministry of the Interior and Safety and National Forensic Service built an AI deepfake-detection model from a 268-team competition, reaching about 92% verification-test accuracy, and deployed it to screen election content ahead of the June 2026 local elections.

388 cases
Illegal deepfake election cases, prior legislative-election cycle
10,510 cases
Illegal deepfake election cases, 2024 presidential-election period
268 teams
Teams in detection-model competition (December 2025)
1,077 participants
Participants in detection-model competition (December 2025)
~92 %
Verification-test detection accuracy
South Korea's AI Deepfake Detection Model for the 2026 Local Elections

Details

Maturity
Scaling
Promoter
Ministry of the Interior and Safety & National Forensic Service
Period
2025-2026
Keywords
elections, deepfake detection, digital forensics, public safety

Context

South Korea saw a surge in illegal AI-generated and deepfake election material, from 388 cases in a prior legislative-election cycle to 10,510 cases during the 2024 presidential-election period, per Ministry of the Interior and Safety data.

Objectives

The Ministry of the Interior and Safety and the National Forensic Service set out to build a dedicated detection model ahead of the 9th nationwide simultaneous local elections, held 3 June 2026.

Activities

An 'AI Detection Model Competition for Deepfake Crime Response' held in December 2025 drew 268 teams and 1,077 participants; the five best-performing models were integrated into one system combining global video-flow analysis with local analysis of facial features, lip movement and skin texture.

Results

The combined system reached about 92% detection accuracy in verification testing, and a rapid-assessment support system now routes flagged content to the National Election Commission and investigative agencies during the election period.

Conclusions

Published figures describe only pre-election verification-test accuracy and prior election-cycle case counts, with no independent post-election field-accuracy audit of the June 2026 deployment found in public reporting; officials plan to extend the same infrastructure to other AI-enabled crimes including sexual exploitation material, defamation and financial fraud.

Implementation
Indicative cost
Medium (€50k–€500k)
Time to results
Short (< 1 year)
Staffing & skills
Ministry of the Interior and Safety, National Forensic Service, 268 competition teams / 1,077 participants who built candidate detection models, National Election Commission and investigative agencies receiving flagged content

Conditions for success

  • an open competition format to source and select the five best-performing detection models
  • combining global video-flow analysis with local facial/lip/skin-texture analysis
  • a rapid-assessment routing pipeline to election and investigative authorities

Common failure modes

  • only pre-election verification-test accuracy (~92%) is published; no independent post-election field-accuracy audit was found
  • no disclosed false-positive rate or appeals process

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful