Securing the collective intelligence: a comprehensive review of federated learning security attacks and defensive strategies

Federated learning holds significant potential as a collaborative machine learning technique, allowing multiple entities to work together on a collective model without the need to exchange data. However, due to the distribution of data across multiple devices, federated learning becomes susceptible to a range of attacks. This paper provides an extensive examination of the different forms of attacks that can target federated learning systems. The attacks discussed include data poisoning attacks, model poisoning attacks, backdoor attacks, Byzantine attacks, membership inference attacks, model inversion attacks, etc. Each attack is examined in detail, with examples from the literature provided. Additionally, potential countermeasures to defend against these attacks are explored. The objective of this review is to provide an in-depth survey of the current landscape in federated learning attacks and corresponding defense mechanisms.