Table 1. AdvFaceGAN’s training process.

Input:source face x, target face y. Output:The parameters of θD and θG after convergence.
1     Initialize parameters of θD andθG,load Dataset
2     while Ladv_target > Ladv_source do:
3     for c in epoch_size do:
4             Get batch with x,y from Dataset
5             ϵ = G(x, y)
6             $\tilde{x}=x+ϵ$
7             $⌢x=\sigma x+\left(1-\sigma \right)\tilde{x}$
8             $L_D=\frac{1}{\mathrm{m}}\left[{\sum }_{i=1}^{m}D\left({\tilde{x}}_i\right)\right]-\frac{1}{\mathrm{m}}\left[{\sum }_{i=1}^{m}D\left(x_i\right)\right]+{\lambda }_{gp}\frac{1}{\mathrm{m}}\left[{\sum }_{i=1}^m{\left({∥{\nabla }_{⌢x_i}D\left(⌢x_i\right)∥}_2-1\right)}^2\right]$
9             θD = Adam(∇DLD, θD, β1, β2, lr, weight_decay)
10             $L_{gan}=-\frac{1}{m}\left[{\sum }_{i=1}^{m}D\left(\tilde{x}\right)\right]$
11             $L_{adv\text{\_}source}=\frac{1}{T\ast m}{\sum }_{i=1}^m{\sum }_{j=1}^T\left(1-cos\left[f_j\left({\tilde{x}}_i\right),f_j\left(x_i\right)\right]\right)$
12             $L_{adv\text{\_}target}=\frac{1}{T\ast m}{\sum }_{i=1}^m{\sum }_{j=1}^T\left(1-cos\left[f_j\left({\tilde{x}}_i\right),f_j\left(y_i\right)\right]\right)$
13             Ladv = η∗Ladv_source + Ladv_target
14             $L_{pert}=\frac{1}{m}{\sum }_{i=1}^{m}max\left(\varepsilon ,{∥{ϵ}_i∥}_2\right)$
15             $L_{st}=\frac{1}{m}{\sum }_{i=1}^{m}max\left(1-\zeta ,1-ssim\left(x_i,{\tilde{x}}_i\right)\right)$
16             LG = Lgan + λpertLpert + λadvLadv + λstLst
17             θG = Adam(∇GLG, θG, β1, β2, lr, weight_decay)
18     end for
19     save θD and θG
20 end while