Police Deepfake and Synthetic Media Detection

A deepfake detection tool takes a video file and returns a judgement on whether it was generated or manipulated. It is a classifier, not a proof: it produces a probability, and the probability is wrong some of the time in both directions.

HOW DETECTION ACTUALLY WORKS

Three families of signal are used in the systems documented so far. The first is artefacts of face generation and replacement, the seams left where a synthetic face is composited onto a real head. The second is physiological: a real face recorded on video carries faint periodic colour changes from blood flow, and blink patterns that generators reproduce imperfectly. The third is file-level, examining metadata and encoding history for traces of the tools used.

A separate and increasingly important step is attribution rather than detection: working out which generative program produced a file, and which source images it was built from, so that a distribution chain can be reconstructed.

THE KOREAN SYSTEM IS THE CLEAREST DOCUMENTED CASE

The Korean National Police Agency's National Office of Investigation put a detection tool into use in March 2024. Korean reporting of the announcement gives a training corpus of 5.2 million data points from 5,400 people, including one million Korean samples and 130,000 from other Asian populations, and an analysis time of five to ten minutes per video depending on quality and length.

The deliberate weighting of the corpus towards Korean faces is worth noting. Detection accuracy, like face recognition accuracy, varies by the population a model was trained on, and building a national tool on a national corpus is a reasonable response to that.

WHAT IT HAS BEEN USED FOR

By August 2026 the tool had been applied in 1,636 investigations, including 72 election crime cases during 2026. Korean police reported 419 arrests for deepfake offences between January and July 2026, with 17 people detained. Those arrests are the product of whole investigations and not a measure of the software's performance.

The agency budgeted 27 billion won for 2025 within a programme of 91 billion won running to 2027 to develop detection of both synthetic video and synthetic voice.

ACCURACY IS THE WHOLE PROBLEM

The reported detection rate is about 80 per cent. Officials said explicitly at launch that because it is not complete, results would guide the direction of an investigation rather than serve as evidence. That is the right framing and an unusually candid one.

It also sets the limit of the technology. At 80 per cent, roughly one judgement in five is wrong, and the two failure modes are not equivalent. A false negative lets a fabricated video pass as real. A false positive tells investigators that a genuine recording, which may be evidence of an actual offence against an actual person, is a fabrication. In the sexual-image cases that dominate the Korean caseload, both errors fall on complainants.

Detection is also adversarial in a way that most forensic tools are not. Generators improve continuously, and a classifier trained on last year's outputs degrades against this year's. A published accuracy figure is a measurement of one moment, not a property of the tool.

THE CASE FOR IT

The alternative to a classifier is not certainty; it is a specialist examining every file by eye, which does not scale to the volume of material now in circulation. Used as triage, a tool that sorts a large seizure into probably-real and probably-synthetic lets scarce forensic expertise go where it is most needed, and the Korean model routes ambiguous results to human examiners rather than accepting the software's answer. Where synthetic sexual imagery is generated at volume by cheap consumer tools, no purely manual response is realistic.

WHO OVERSEES IT

No published oversight framework for police deepfake detection was identified. There is no known independent evaluation of any deployed system's accuracy, no published standard for when a detection result may be put before a court, and no established disclosure obligation requiring a defendant to be told that a classifier was used to direct the investigation against them.

WHAT IS NOT KNOWN

Whether the reported 80 per cent holds against current generators is unknown, as is whether accuracy varies across skin tone, age or video quality. Whether the Korean tool was built wholly in-house is unresolved: police describe it as domestically developed and name no supplier, while a commercial vendor separately announced a detection solution developed in collaboration with the agency, and no source read establishes whether these are the same system. Deployments outside South Korea are not well documented.