Algorithm 1 System model of FL
Input: Set of connected clients C , E is the number of global iteration , K is the number of local epoch, $\eta$ is the learning rate , b is the size of local mini batch , r is rate of client selection
Output: Global model W
1: procedure Server($E,K$)	▹ Central Server execute
2:   $W_0,k\leftarrow Initialization$	▹ Initialize global model, constant K
3:   for global iteration $t\le E$ do
4:    $N\leftarrow max\left(\right|C|\times r,1)$
5:    $S_t\leftarrow$ ClientSelection (C, N)	▹ Select client for train
6:    for each client $c_x\in S_t$ in parallel do
7:     Broadcast $W_t,K$ to client $c_x$
8:     $W_t^x\leftarrow$ ClientUpdate $(c_x,K,W_t)$	▹ aggregate model
9:    end for
10:    Update global model $W_{t+1}\leftarrow {\sum }_{x=1}^{|S_t|}\frac{|D_x|}{D_t}{W}_t^x$
11   end for
12: end procedure
13:
14:
15: procedure clientupdate($c_x,K,W_t$)
16:   B← split local data $D_x$ into batches of size b
17:   Replace local model $w_t^x$←$W_t$
18:   for local epoch e≤K do
19:    for batch data b∈B do
20:     $w_{t,e}^x\leftarrow w_{t,e}^x-\eta ▿F_x(w_{t,e}^x,b)$	▹ mini-batch SGD
21:    end for
22:   end for
23:   Upload local update result $W_t^x\leftarrow w_{t,K}^x$
24: end procedure