Byzantine-Resilient Federated AI with Low Comm. Cost

Federated learning is attractive for privacy-sensitive settings, but current approaches often require transmitting high-dimensional gradients, creating major bandwidth and compute bottlenecks. At the same time, even a few malicious or faulty participants can impair convergence or derail training. These limitations restrict deployment in real-world multi-party environments with sensitive data and heterogeneous infrastructure.

The invention enables distributed model training through highly compressed scalar-based updates instead of full gradient exchange. Shared random seeds allow clients and server to reconstruct update directions locally, drastically reducing communication overhead. Robust aggregation protects training against Byzantine participants, while preserving model performance. The approach has been validated on standard machine learning tasks and large language model fine-tuning, demonstrating broad applicability in privacy-sensitive, resource-constrained settings.