https://cis.temple.edu/~wu/research/publications/Publication_files/J-CT-2024-Balancing%20Privacy%20and%20Accuracy%20using%20Siginificant%20Gradient%20Protection%20in%20Federated%20Learning.pdf
Similar to [34], each client has one or more lines for training or testing. We train a recurrent neural network (RNN) model for predicting the next character. RNN takes a sequence of 80 characters as input and consists of an embedding layer (80 × 8), two LSTM layers (80 × 256), and a dense layer (80 × 90).