Qora-FL: Quorum-Oriented Robust Aggregation for Federated Learning

Description:This preprint presents Qora-FL, a high-performance library for Byzantine-tolerant aggregation in federated learning. Qora-FL implements coordinate-wise trimmed mean, coordinate-wise median, and Krum selection, with optional Q16.16 fixed-point arithmetic for deterministic aggregation. The system was derived from and validated by a 181-day autonomous IoT deployment (QRES) operating under persistent 30% Byzantine participation, and adapts these consensus mechanisms to server-side federated learning as a drop-in strategy for the Flower framework. Experimental evaluation demonstrates sub-10ms aggregation latency for models with 100K parameters, stable convergence under 30% label-flipping attacks on MNIST, and the effectiveness of a deviation-derived reputation mechanism for filtering adversarial clients over time.