Federated Learning Architectures and Communication-Efficient Optimisation for Privacy-Preserving Distributed AI Systems

Abstract

Federated learning (FL) trains a shared model across data holders that cannot pool their records, but deployments remain bounded by three coupled costs: uplink traffic from repeated model exchange, accuracy loss under statistically heterogeneous clients, and the information that updates still leak. These are usually attacked in isolation, and the resulting designs interact poorly. Aggressive sparsification concentrates the whole differential-privacy noise budget onto the few coordinates that survive, so a compressor tuned without reference to the noise scale erodes the utility the privacy mechanism was calibrated to preserve. We present FedHCP, a three-tier architecture that budgets compression and privacy jointly. Clients apply error-compensated top-ρ sparsification, clip the retained update, and perturb only the transmitted support; edge aggregators combine masked updates so no single server observes an individual contribution; a cloud coordinator aggregates cluster deltas with momentum. We derive a closed-form retention ratio minimising an upper bound on per-round distortion as an explicit function of the noise multiplier, and drive an online controller with it. Under a simulation protocol on Dirichlet-partitioned CIFAR-10 and on FEMNIST, the design matches uncompressed federated averaging while transmitting about 37× fewer uplink bytes, and holds 78.9% CIFAR-10 accuracy at (ε=8,δ=10−5). Results are reported as a design-level evaluation of the stated protocol, not as deployment measurements.