Revealing byzantine attackers through smart re-clustering in federated learning

A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising:

• detecting an attack in a federated learning system that includes clusters of clients that are configured to perform federated learning;
• generating a cluster score for each of the clusters, wherein the cluster score identifies which of the clusters are suspect clusters and which of the clusters are fair clusters;
• identifying the clients in the fair clusters as fair clients;
• identifying clients in the suspect clusters as suspect clients;
• repeating until a malicious client is identified:
• reclustering the clients into new clusters;
• determining a new score for each of the new clusters, wherein the new score identifies which of the new clusters are new suspect clusters and which are new fair clusters;
• labelling all suspect clients in the new fair clusters as fair clusters without changing labels of clients in the new fair clusters that are already labeled as fair;
• and
• labelling a client in one of the new suspect clusters as the malicious client when all other clients in that new suspect cluster are labeled as fair clients.