EarlyBirdFL: Leveraging Early Bird Ticket Networks for Enhanced Personalized Learning

Federated learning (FL) is revolutionizing mobile computing and IoT development by enhancing data privacy. However, restricted computational and communication resources and the statistical variability of data stored on devices present substantial obstacles to ongoing progress in FL. We introduce EarlyBirdFL, a novel FL framework that leverages an Early-Bird Ticket-inspired pruning and masking technique for efficient training and communication in federated settings. EarlyBirdFL enables each client to achieve fast local training by identifying efficient subnetworks early in the training process, communicating only these pruned networks between the server and the client. Unlike classical personalized FL, in which the client-side model learns differences, EarlyBirdFL allows each client to identify these efficient subnetworks using a mask metric quickly. Experimental results demonstrate that EarlyBirdFL outperforms traditional computation time and accuracy methods, achieving a 1.53-4.98 times speedup and 1.01-1.15 times higher accuracy. Furthermore, EarlyBirdFL remains stable even when its parameters are adjusted and performs well in different non-IID environments, maintaining or surpassing the performance of other methods. This approach combines elements of early efficient subnetwork identification, pruning, masking, and personalized federated learning to address the unique challenges of FL.