CyBeR-0 introduces a novel zero-order optimization method that protects federated learning systems from Byzantine attacks while significantly reducing communication costs.

• Provides robust aggregation for heterogeneous client data
• Maintains stable performance while transmitting only a few scalars
• Offers convergence guarantees for non-convex objectives
• Successfully tested on standard learning tasks and LLM fine-tuning

This research is critical for security in distributed learning environments where malicious actors might compromise client nodes, enabling safer deployment of federated learning in sensitive applications.

Byzantine-Resilient Zero-Order Optimization for Communication-Efficient Heterogeneous Federated Learning