This week

• [x] FL vs Centralized (NYC and LDCC)
• [x] Population vs Demographic (Synthetic, LDCC)
• [x] FL vs FL with DP (LDCC)

Google Colaboratory

In order to improve the model performance, i have tested out different combinations of the hyperparameters.

The only changes which have led to a little improvements are the insertion of a dropout of 0.3 in the GRU layer.

At the moment, the best model settings I have been able to find are:

• N = 17
• NUM_EPOCHS = 5
• BATCH_SIZE = 16
• embedding_dim = 256
• rnn_units = 256
• learning_rate=0.004
• learning_rate_client=0.04
• drop_rate = 0.3

Architecture:

Embedding -> GRU -> Drop -> Dense