Differential privacy in peer-to-peer federated learning

Abstract

Neural networks have become tremendously successful in recent times due to larger computing power and availability of tagged datasets for various applications. Training these networks is computationally demanding and often requires proprietary datasets to yield usable insights. In order to incentivise stakeholders to share their datasets in order to build stronger neural networks and protect their privacy interests, it is important to implement differential privacy mechanisms during the training of neural networks to protect against attacks that might expose their data to malicious agents. The objective of this project is to study the effectiveness of differential privacy implementation on peer-to-peer federated learning in protecting proprietary data from exposure.