Privacy-Preserving Federated Graph Neural Network Learning on Non-IID Graph Data

Algorithm 2

The federated GAT framework.

Input:: the set of clients

  G_i: the local subgraph hold by c_i

  U_k: the public nodes shared among C

Output: the node embeddings of G_i

 1: LOCAL CLIENTS:

 2: for each client c_i ∈ C do

 3:  Compute the GAT model embedding

 4:  Generate the public nodes’ embeddings X_i of U_k

    from the intermediate H_i:

 5:    

 6:  Upload X_i to the server

 7: end for

 8:

 9: while not converge do

 10:  SERVER:

 11:  for each i ∈ kdo

 12:   for each do

 13:    Align X_j into c_i’s space: X_ji = W_jiX_j

 14:   end for

 15:   Aggregate all the aligned embeddings with X_i

 16:    

 17:   distribute to client c_i for local update

 18:  end for

 19:

 20:  LOCAL CLIENTS:

 21:  for each client c_i ∈ C do

 22:   Take as new input weights

 23:   Compute the GAT-model embedding with loss L_new

 24:  end for

 25: end while

 26: return the node embeddings of G_i