Brief

Researchers have developed Giskard, a new protocol designed to enhance security and efficiency in large-scale decentralized learning. Giskard addresses the challenge of simultaneously maintaining data confidentiality and defending against Byzantine (malicious or faulty) participants. The protocol organizes participants into a tree of committees, enabling a more scalable approach to aggregation compared to existing methods that require all-to-all communication or heavily burden a small subset of nodes. Experiments with up to one million participants demonstrate Giskard's ability to reduce communication complexity while maintaining model utility even with a significant proportion of Byzantine parties. AI