Byzantine-robust aggregation is a standard defense against malicious updates in federated learning. Existing filtering rules work by comparing clients' parameter vectors, which requires the server to see every client's update.
The new arXiv paper, LiBRA, proposes an alternative: use logit fingerprints for Byzantine-robust client selection. This shifts the comparison away from full parameter vectors, addressing a key limitation of current approaches.
The abstract is brief and does not include experimental results or implementation details. What is clear is the motivation: reduce reliance on raw parameter vectors when deciding which clients are trustworthy.