FedMID: A Data-Free Method for Using Intermediate Outputs as a Defense Mechanism Against Poisoning Attacks in Federated Learning

arXiv:2404.11905v1 Announce Type: new
Abstract: Federated learning combines local updates from clients to produce a global model, which is susceptible to poisoning attacks. Most previous defense strategies relied on vectors derived from projections of local updates on a Euclidean space; however, these methods fail to accurately represent the functionality and structure of local models, resulting in inconsistent performance. Here, we present a new paradigm to defend against poisoning attacks in federated learning using functional mappings of local models based on …

abstract arxiv attacks cs.cr cs.lg data defense federated learning free global however intermediate poisoning attacks space strategies type updates vectors