Miyaji Laboratory

Federated Learning (FL) is a method that enables global optimization while keeping client training data confidential. Each client trains a model, the server collects the model weights, constructs an integrated model, and distributes the weights. Split Computing (SC) is a method that reduces processing load while keeping client training and inference data confidential by splitting processing between clients and the server. Split-GP has been proposed as a method combining FL and SC. In Split-GP, each client uses the first half of the model as their client model, while the server uses the second half as the server model. During training, each client extracts features and sends intermediate features to the server. At
this time, the client also calculates and stores its local gradients. The server computes the loss from the received intermediate features and returns gradients to the clients. The server updates its model based on these gradients. Clients perform backpropagation using the received gradients combined with their local gradients to update their client models. During inference, each client first performs inference using its own model. If the inference result is uncertain, it offloads the inference to the server model. However, this approach employs Multi-Exit, which allows inference to terminate at shallow layers on the client, potentially enabling higher individualized performance. Furthermore, within the same Federated Learning (FL) framework, malicious clients could potentially contaminate other clients' models. This research constructs client-specific models that are specialized for individualization performance for each client, free from influence by other clients, by not performing backpropagation of gradients from the server on each client. The server constructs a generalized model by integrating the server models created by each client. At this time, we verify communication volume and accuracy under heterogeneous data conditions.