AI and machine learning (ML) have significantly impacted our daily lives, including financial services. For example, they are used in fraud detection, predicting loan defaulters, portfolio management, stock price predictions for algorithmic trading, personalised customer support through chatbots, and more.

Consider a specific scenario: a company wants to create a module for predicting loan defaulters and offer it as a service to different banks. The company aims to enhance customer experiences and refine the underlying machine learning models. To achieve this, the company needs to gather data from client banks to improve the models. However, due to legal and privacy concerns, banks may be hesitant to share their data. Such situations have led to the concept of federated learning (FL).

Challenges with FL in Fintech
Such a decentralised approach in the financial sector is the way to advance. However, it poses challenges. Though the clients are not sharing data, there is the possibility of model inversion attacks, which may predict the presence of specific individuals. It may be problematic if the data is about loan applicants, and the applicant would not be the financial institution’s customer who leaked its data. Another challenge is: Is such a model fair across demographics? From a bank’s perspective, it may prefer to free-ride on other banks’ contributions unless offered incentives.

IIITH’s Contributions
At the machine learning laboratory in IIITH, our group has looked at all three challenges in adapting AI through FL in financial services and proposed innovative solutions combining machine learning and game theory. We have developed techniques that optimally learn in FL with minimal compromise on privacy. We have used incentive engineering concepts to ensure that all the participants receive fair compensation for their contributions. We proposed heuristics on combining different client models to ensure that the aggregated model is fair across demographics. Though we derive motivation from financial domains, the techniques we developed are general and can be adapted to multiple other domains as well.