DG-FedReuse Boosts Federated Learning Efficiency with Cached Updates.

Federated learning (FL) involves repeated local model optimization and subsequent transmission of updates from clients to a central server, which can be communication-intensive. A new mechanism, DG-FedReuse, aims to mitigate this by allowing selected clients to reuse age-decayed cached updates. This reuse is conditionally gated: it occurs only when a stochastic head-gradient discrepancy proxy remains below a round-dependent threshold. The system incorporates constraints such as a hard cache-age limit and a minimum quota for fresh client updates to ensure model freshness and prevent excessive staleness. Additionally, fresh updates utilize an adaptive per-tensor Top-K numerical-field representation to further reduce uplink data. Simulations across six image-classification datasets with 50 virtual clients and Dirichlet label heterogeneity demonstrated significant communication savings. DG-FedReuse achieved 83.36-85.42% modeled update-data-field uplink saving, substantially outperforming matched Top-K FedAvg's 76.88% saving. While accuracy differences ranged from -5.29 to -0.14 percentage points compared to matched FedAvg, the primary benefit lies in the drastic reduction of communication overhead. The study emphasizes that communication conclusions depend on the accounting boundary, noting that a symmetric dense-model-downlink sensitivity reduces the headline saving but still shows incremental gains over Top-K FedAvg.