Kobe University, National Information and Communication Research Organization, and Digital Risk Management Eltes collaborate with Chiba Bank, Mitsubishi UFJ Bank, Chugoku Bank, Sumitomo Mitsui Trust Bank, and Iyo Bank to use privacy protection federated learning technology (* 1) for fraud. A demonstration experiment of remittance detection was conducted.
According to Kobe University, the demonstration experiment was divided into the detection of damaged transactions and the detection of fraudulent accounts.When the detection accuracy of the damage transaction was compared between the detection accuracy of the normal machine learning (* 2) model and the detection accuracy of the federated learning model, the detection accuracy improved with the introduction of the federated learning model and achieved the target of 80% or more. ..In addition, we found fraudulent transactions that could not be detected by the individual learning model.
In the detection of fraudulent accounts, as a result of comparing the accuracy with a hybrid model that combines an individual learning model, an individual learning model, and a federated learning model, the accuracy of the hybrid model was high, and a detection rate of 80% or more was achieved.It was also found that it can be detected about 20 to 50 weeks earlier than the fraudulent account freeze in the actual data.
Kobe University and others have been selected for the science and technology promotion organization adoption project in 2022, and will continue to improve the detection function and implement the system in the future.
* 1 Privacy protection Federated learning technology Technology for machine learning in a distributed state without aggregating data in one place
* 2 Machine learning A technology in which a computer discovers certain rules from certain data and makes inferences and predictions for unknown data based on those rules.