Secure Federated Analytics on Vertically Partitioned Data

点击翻译按钮获取中文摘要

The goal of this project is to improve the accuracy of algorithms to detect financial crime by analysing transaction information, while at the same time ensuring that the privacy of individuals and organisations is protected. The scale of the financial crime is vast: the UN estimates that US$800-2000bn is laundered each year, representing 2-5% of global GDP. A more successful approach for detection can reduce this amount as well as the expense and inconvenience caused when genuine account transactions are flagged as potentially fraudulent.Accuracy can be improved through the collaboration of different financial institutions to piece together the information they hold about the accounts and transactions involved, building up a clearer picture of the characteristics and patterns that indicate fraud.However linking together information also results in richer profiles for each transaction and each account. This can increase privacy risk by providing more possibilities to recognise individuals in the dataset and reveal their sensitive information.The project seeks to develop an approach to carry out analysis on the data held across different financial institutions to improve the accuracy of crime detection without collecting and centralising the data one place. A federated learning approach is used to derive predictive features from the data and to train a machine learning model without sharing confidential individual records. Together the organisations are able to produce a high-accuracy model which can then be deployed to monitor and flag potentially problematic transactions.The project seeks to enable collaborative analysis while preventing confidential information being shared across financial institutions and to limit any information that can be learned about innocent individuals from deployment of the machine learning model.