Study on attacks against federated learning

Abstract

Increasingly strict data privacy laws have seen many companies that are taking advantage of big data flock over from simple collaborative learning systems to federated learning systems which promise the reservation of data privacy. However, due to the collaborative and distributed nature of federated learning, the resulting trained model will still be very exposed and vulnerable to many other kinds of attacks by malicious or compromised participants. In this project, we aim to study the various attack and defence methodologies that can be deployed in federated learning by implementing them using an existing open-sourced federated learning implementation as the base code. Such studies will allow us to understand the threats to the federated learning process and subsequently how to mitigate or even prevent any detrimental effects coming from the threat actors. The focus for this project will be on the distributed backdoor attack methodology and the PDGAN defence methodology.