KAIST Successfully Upgrades Software
Accurately Detects Tampering Using Color and Frequency Information in Videos
[Asia Economy Reporter Kim Bong-su] Domestic researchers have developed a technology that accurately detects image forgery by analyzing color and frequency information within videos using artificial intelligence (AI) for the first time in the world.
The Korea Advanced Institute of Science and Technology (KAIST) announced on the 13th that the research team led by Professor Lee Heung-kyu of the Department of Computer Science developed version 2.1 of 'KaiCatch,' an image forgery detection software, which significantly improves the precision and accuracy of video image forgery detection by employing a new AI architecture and learning methodology, as well as using advanced transformed image videos that are difficult to obtain in laboratory environments. This version can also detect video editing manipulations.
The KaiCatch software consists of two AI engines: the 'Anomaly Type Analysis Engine' and the 'Anomaly Region Estimation Engine.' The 'Anomaly Type Analysis Engine' defines essential variations such as blurring, noise, size changes, contrast changes, morphing, and resampling, and detects them. The 'Anomaly Region Estimation Engine' detects image splicing, cutting and pasting, copy-pasting, and copy-moving. The newly developed technology focuses on the 'Anomaly Region Estimation Engine.' Previously, anomaly region detection was performed in grayscale, which had low representational power for analysis signals and resulted in many detection errors, making it difficult to determine forgery. The newly developed technology utilizes both color and frequency information, greatly improving precision and recall, and expresses the manipulated regions in color scale, enabling clearer identification not only of anomalies in the region but also of forgery.
In this study, the research team proposed for the first time in academia an approach that analyzes both color and frequency information together to examine traces generated during video creation and compression. To implement this methodology, they developed the 'Compression Artifact Tracing Network (CAT-Net),' which directly accepts frequency information as input within a single segmentation network, marking the first such development in academia. They demonstrated that CAT-Net significantly outperforms existing methods in detection performance. Compared to previously proposed techniques, the developed technology excels particularly in evaluation metrics such as the F1 score and average precision, which distinguish between original and manipulated versions, greatly enhancing real-world forgery detection capabilities.
Regarding video editing manipulations, the team noted that editing changes such as frame deletion and addition frequently occur in CCTV videos. Accordingly, the capability to detect such video editing manipulations has also been incorporated into KaiCatch version 2.1.
Professor Lee’s team previously developed KaiCatch as an Android app embedded in smartphones in March last year. Since then, they have received over 900 forgery analysis requests via the KaiCatch app and more than 60 individual detailed forgery analysis requests. However, the high false detection rate resulted in actual detection precision being much lower than theoretical values, making clear technical judgments on forgery or manipulation often impossible.
Professor Lee stated, "The newly developed KaiCatch 2.1 uses a new network structure called CAT-Net, a learning methodology, and advanced technology called ‘simultaneous analysis of color and frequency domain distortion traces’ to improve precision and enable clearer judgments. We hope that cases where it is difficult to determine video forgery will significantly decrease in the future."
The research team focused on MP4 file format for videos and JPEG images for video images, as these formats are widely used by the general public. For video images, the emphasis was on detecting artificially generated JPEG compression subtle signals left during video editing manipulations, concentrating on identifying forgery and forged regions. For videos, the system detects cases where specific frames are deleted or inserted, or frames are partially edited and then recompressed. Given the recent increase in disputes over CCTV video editing, this technology is expected to be highly beneficial.
Currently, the KaiCatch software can be downloaded and installed by searching for ‘KaiCatch’ on the Google Play Store for Android-based smartphones. Users can simply upload video images to KaiCatch to test for forgery.
The research results were published online on the 25th of last month in the 'International Journal of Computer Vision (IF 7.410),' an international journal in the field of computer vision published by Springer Nature.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
