GAN FOR DATA SYNTHESIS WHILE MINIMIZING PREDICTIVE POWER LOSS

Abstract

In the modern financial environment, the ability to leverage data for developing intelligent risk detection methods is a huge ethical challenge. However, this does not make it less of a necessity. That is something I need to solve for my research as well. The quality and quantity of the data I would be able to employ for my research hinges on how good would be my anonymization proposal. Anonymization is the process of transforming datasets to prevent identification of individuals or organizations. It has become of paramount importance in order to ensure privacy and compliance in experiments involving sensitive financial information. As regulatory frameworks such as the General Data Protection Regulation (GDPR) in the European Union and similar local laws worldwide are on the rise, ethical approaches to anonymization are no longer optional but essential to achieve legal compliance and to conduct responsible research. Advances in machine learning and big data analytics have increased the likelihood of re-identification, especially when anonymized data is combined with external datasets [1]. This raises critical questions about whether traditional anonymization techniques remain sufficient in high-risk financial contexts and can be still used to full extent. However, it is known that over-anonymization will degrade data quality, sometimes severely, limiting the ability of machine learning algorithms to detect necessary patterns in order to locate fraudulent transactions efficiently. Ethical practice requires balancing privacy protections with the need for accurate and efficient models. Finding this balance is proving to be even more challenging as the demand for the data gets bigger. Removing key values will unintentionally bias models, particularly in fraud detection, potentially worsening their performance as well as their interpretability.