Coded Computing for Low-Latency Federated Learning Over Wireless Edge Networks

Federated learning enables training a global model from data located at the client nodes, without sharing and moving to centralized server. Performance of federated in multi-access edge computing (MEC) network suffers slow convergence due heterogeneity stochastic fluctuations compute power communication link qualities across clients. We propose novel coded framework, CodedFedL, that injects structured coding redundancy into for mitigating stragglers speeding up procedure. CodedFedL non-linear by efficiently exploiting distributed kernel embedding via random Fourier features transforms task computationally favourable linear regression. Furthermore, clients generate local parity datasets over their datasets, while server combines them obtain dataset. Gradient dataset compensates straggling gradients during training, thereby speeds convergence. For minimizing epoch deadline time MEC server, we provide tractable approach finding amount number points processes statistical properties as well delays. also characterize leakage privacy when share with analyze rate iteration complexity under simplifying assumptions, treating gradient descent algorithm. conduct numerical experiments using practical parameters benchmark where overall $15\times$ comparison schemes.