An Architecture for Resilient Federated Learning through Parallel Recognition

In federated learning, non-independent and identically distributed (non-IID) local datasets lead to accuracy loss compared to homogeneous distribution of datasets. In this paper, we propose an architecture for improving accuracy and offering resilience through federation utilizing non-IID datasets. The proposed architecture performs parallel recognition employing triplication of AI processors with different intelligence. Experimental results demonstrate that the proposed architecture improves accuracy by 2.3% compared to accuracy of a single AI processor on average and guarantees resilience.