Catching 'Deepfake' with Korean AI Technology

Practical Software for Detecting Photo Forgery and Manipulation
Opening the Way for the First Full-Scale Commercialization of Photo Forgery Detection Technology in Korea

[Asia Economy Reporter Junho Hwang] An artificial intelligence that detects various forged and altered photos and videos, including 'deepfakes'?fake photos or videos created using artificial intelligence?has been developed by a domestic research team. This is the first AI in Korea that transforms forgery detection technology, which had remained at the paper level, into practical software.

The research team led by Professor Heungkyu Lee of the Department of Computer Science at the Korea Advanced Institute of Science and Technology (KAIST) announced on the 3rd that they have developed practical software called 'KaiCatch' that detects whether digital photos have been altered using artificial neural networks.

KaiCatch Detects Wide-ranging and Diverse Alterations

Detection Results by Essential Variant Attack

KaiCatch is an AI specialized in detecting a broad range of alteration types rather than focusing on specific alterations. It overcomes the limitations of existing technologies that gather individual algorithms for detecting specific alterations, enabling organic detection of various modifications.

The research team defined essential variations that always occur in common and frequently occurring alterations such as cutting and pasting, copying and pasting, erasing, changes in object size and movement within images, and retouching. They then made it possible to comprehensively detect these, allowing the system to determine whether an alteration has occurred even if the specific type of alteration cannot be identified.

To develop KaiCatch, the team utilized over 300,000 circulated image data collected from June 2015 through a 'Digital Image Forgery Identification Web Service' targeting the general public, forensic video data based on feature and neural network methods, and a large amount of experimental video materials for deepfake and steganalysis.

They also employed technology that comprehensively processes JPEG images based on more than 50 standard quantization tables and over 1,000 non-standard quantization tables, including uncompressed and lossless compressed formats such as BMP, TIF, TIFF, and PNG.

First in Korea, Second in the World: AI for Forgery Detection

Photo Editing Manipulation Detection Results

The research team stated, "It is significant that we have brought photo and video forgery detection technology, which had only remained at the paper publication level, to a practical stage for the first time in Korea and the second time worldwide."

Professor Lee said, "KaiCatch incorporates video forensic technology that combines pixel-level alteration detection, which commonly occurs across various modifications, with artificial intelligence technology. This technology shows excellent performance in determining whether a given digital photo has been altered in arbitrary environments."

He added, "In the future, we plan to add extensive detection functions for advanced features of various editing tools and to develop the currently laboratory-level deepfake detection engine and general video alteration detection engines to practical levels and integrate them into KaiCatch."

This content was produced with the assistance of AI translation services.