Privacy-Preserving Federated Learning for Building Energy Forecasting: A Differential Privacy Analysis of Aggregation Strategies
This work trains a CNN-LSTM forecasting model under three FL aggregation strategies under three DP-FL aggregation strategies, and advances per-sample gradient clipping in differentially private stochastic gradient descent (DP-SGD), which constrains heterogeneous client updates as the most plausible mechanism.
J. .. Ameh, Abayomi Otebolaku, Augustine Ikpehai
· Energies · 0 citations