SplitFed Learning

A hybrid distributed collaborative machine learning algorithm and dual-server split-network architecture (SplitFed / SFL; Thapa, Chamikara, Camtepe, and Sun, AAAI 2022) that partitions a full neural network W at a cut layer into a client-side sub-network W^C (of size beta |W|) and a server-side sub-network W^S while training all K clients in parallel rather than sequentially. In each global epoch, clients execute forward propagation on W_{k,t}^C in parallel (optionally injecting calibrated Laplacian PixelDP noise A_{k,i}^P = A_{k,i} + Lap(Delta I_{p,q}^A / eps') before the cut layer) and transmit cut-layer activations ('smashed data' A_{k,t}) to a Main Server; the Main Server computes server-side forward-backward passes (in parallel per client followed by sample-weighted server model averaging W_{t+1}^S = W_t^S - eta sum_{k=1}^K (n_k / n) nabla ell_k(W_t^S; A_t^S) in SFLv1, or sequentially across randomly ordered clients' smashed data without server-model averaging in SFLv2), returns cut-layer gradients dA_{k,t} to each client for client-side backpropagation (with optional per-sample l2 clipping and Gaussian noise for (eps, delta)-DP), and a Fed Server aggregates client-side sub-networks via W_{t+1}^C = sum_{k=1}^K (n_k / n) W_{k,t}^C.