Merge/combine clients with server

i am currently trying to make split learning setup. I have 5 clients and one server.
Each client has own data set (cifar10 divided on five parts))
In the training: each client train network up to cut layer and outputs at the cut layer are sent to the server which completes the rest of the training .I use resnet18

At the end I have clients with accuracy ca 80% :client 1 recognize only plane and car, client 2 recognize only bird and cat etc

Now I need to merge each client with the server, so that each client can recognize whole cifar10(all classes).

Anyone idea how to do that?
I tried to do models aggregation, but nothing is happening. Thanks in advance!