KAIST Launches Korea's First Deepfake Detection Mobile App Service

Development of 'KaiCatch' App Service...Enabling General Public to Detect Digital Content Forgery

[Asia Economy Reporter Kim Bong-su] The Korea Advanced Institute of Science and Technology (KAIST) has launched a neural network-based mobile app service capable of detecting various forged photos, including deepfake videos. It is expected to help prevent damage by allowing the general public to easily utilize it.

On the 30th, KAIST announced that Professor Lee Heung-gyu's research team from the Department of Computer Science developed and started offering 'KaiCatch,' a deepfake and photo forgery detection software based on artificial neural networks, in the form of a mobile app. This is the first service in Korea that enables ordinary people to easily use digital content forgery detection technology. By searching for 'KaiCatch' on the Google Store and downloading the app, users can simply analyze deepfake and image forgery.

Deepfake technology, mainly targeting human faces, can be broadly divided into three types: face swapping, face reenactment, and face attribute manipulation. Among these, face swapping and face reenactment can cause significant social confusion as fake news when used maliciously and can also be exploited for producing pornography, severely infringing on individual rights. Face attribute manipulation can be misused to tamper with video evidence.

The research team developed signal processing and artificial intelligence technologies that apply detection techniques for subtle deformation signal traces and subtle anomaly signal traces in videos to detect deepfakes regardless of type. They detect deepfakes by analyzing subtle deformations in the facial area and anomaly signal traces in potential geometric distortion areas within the face, such as the nose, mouth, and facial contours.

For deepfake detection technology, when a suspicious deepfake video in avi or mp4 format is provided, it is split into individual frames, converted into images for the frames to be analyzed, and then deepfake detection is performed. Normal detection is possible unless the face in the video is too small (resolution below 128×128) or a significant portion of the person's face in the video is cropped. Therefore, by extracting a single frame from the video, converting it into an image, and uploading it to the KaiCatch app, users can easily verify whether it is a deepfake. The analysis result is displayed as a value from 0 to 100 (%), with higher numbers indicating a higher probability of being a deepfake.

Similarly, for photo forgery detection technology, uploading the relevant image to the KaiCatch app allows users to receive forgery analysis results. KaiCatch can process JPEG images based on over 50 standard quantization tables and more than 1,000 non-standard quantization tables, including uncompressed and lossless compressed formats such as BMP, TIF, TIFF, and PNG. When an image suspected of photo forgery is uploaded, two visualized analysis images are generated from the analysis results. If the main areas containing suspicious forgery features show colors significantly different from surrounding areas or if various colors are mixed only in the main areas, ordinary users can easily determine that those areas have been forged.

The development of this mobile forgery detection app is the first of its kind in Korea and is rare even among advanced countries. It was developed to have versatility against various types of alterations by using advanced technologies such as artificial intelligence and subtle anomaly signal trace analysis techniques. In particular, it detects with a high reliability of around 90%, even when unpredictable or unknown alteration techniques are used.

Professor Lee Heung-gyu said, "Although the app service technology was developed to operate only in an Android-based mobile environment, we plan to release an Apple iOS-based app soon, along with English, Chinese, and Japanese versions." He added, "We will also commercialize forgery detection technologies using methods very different from existing detection techniques and add them to KaiCatch to significantly reduce various exceptional cases where detection is not possible."

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