The Ministry of the Interior and Safety (MOIS) announced that the ‘Voice analysis model for voice phishing’ will be used in the voice phishing investigation including voice analysis, from the end of this month.

Until now, the National Forensic Service (NFS) in Korea has used voice analysis models developed in Russia and the U.K. to conduct voice analysis necessary for voice phishing investigations.

However, since the models were based on foreign language data base, there was an issue of accuracy in identifying Korean criminals.

Moreover, the existing models did not have clustering algorithm function. This has been an issue because a voice phishing criminal organization usually works in a group with each person playing a different role such as investigators, prosecutors, etc. and therefore, clustering algorithm is needed to differentiate and identify each person in the group.

Last year, in this context, the Integrated Data Analysis Center (IDAC) under MOIS and the NFS started to develop a model capable of grouping those involved in a crime and made efforts to improve the accuracy of identifying speakers, focusing on arresting voice phishing criminals.

The new model used the latest AI deep learning technology, leveraging more than a million voice data segments in Korean as well as in foreign languages extracted from approximately 6,000 persons at home and abroad.

For Korean language in particular, more than 100,000 voice data segments of ordinary persons and those of voice phishing scammers owned by the NFS were analyzed. In addition, various learning and performance verification processes on such vast amount of data were repeatedly conducted. As a result, optimal algorithm for identifying voice phishing speakers was created.

Following the development of the model, two rounds of accuracy verification was conducted: A total of 660 voice data segments from 150 persons were analyzed in the first round and 12,000 separate voice data segments from 200 persons in the 2nd round taking into diverse circumstances into account.

The performance verification confirmed that the rate of reading to identify a criminal’s voice was improved by about 77%* compared to that of the existing foreign model.
*Upon analyzing 100 criminal voice data segments, the new model was able to identify up to 51 people while the existing model could identify only about 28 people.

IDAC also explained that grouping criminals* has become possible for the first time in the world through the new model.
*Grouping criminals: Identification and clustering of the same persons through serial comparison processes of criminals’ voices for each case
*For instance, in the case above, it is possible to confirm that accomplices ① to ④ belong to the same criminal organization through analysis.

Meanwhile, according to the latest data released by the National Policy Agency (NPA), 156,249 voice phishing cases have occurred in Korea over the last five years. The loss and damage of victims has exceeded 3 trillion KRW, seriously affecting people’s lives.

It is further analyzed that socioeconomic costs for crime prevention are rapidly increasing, as well.

With the successful completion of the model development, IDAC will actively make use of the model in investigating and arresting voice phishing criminals in cooperation with the NFS and the NPA and promote it to be also utilized overseas.

In addition, the NFS will apply the new model in identifying voice phishing scammers from the end of February.

About 10,000 voice data segments of voice phishing criminals in possession of the NFS will be analyzed for criminal organization grouping and interrogation of additional crimes of arrested criminals.

The new model will be also shared with the NPA to improve the speed of the first investigation as well as the arrest rates of voice phishing criminals. Gradually, it will be applied to investigate various voice-related crimes, including impersonating public institutions, deposit-based lease frauds, etc.

Voices of voice phishing criminals analyzed by IDAC’s new model will be posted on the website of the Financial Supervisory Service in order to raise awareness and prevent further voice phishing scams.

With the training sessions and diverse international events to be held later this year, MOIS will promote the excellent performance and expandability of the new model to developing countries intending to learn Korea’s latest voice-based forensic investigation techniques.

Vice Minister Han Chang-seob of MOIS said, “The newly developed voice analysis model for voice phishing is a tangible achievement of the digital platform government, aimed at solving the current social issues through data analysis. MOIS will continue to identify analysis tasks necessary for Korean citizens and utilize the result of such analysis tasks for building a competent data-based government.”

[Source: Integrated Data Analysis Center, Government Innovation Planning Bureau, & Digital Analysis Division, National Forensic Service, MOIS]