Privacy-Preserving Image Template Sharing Using Contrastive Learning

. 2022 May 3;24(5):643. doi: 10.3390/e24050643

Algorithm 1 Adversarial Training Procedure

Input: dataset $D$ and parameter $γ$

Output: $ϕ, ψ, θ_{t}, θ_{a}$

1:
for every epoch do
2:
Sample a minibatch from dataset
3:
Train $ϕ$ using $L_{p r i v a t e - s u p c o n}$ or $L_{p r i v a t e - u n i c o n}$ in Equations (15) and (16)
4:
end for
5:
for every epoch do
6:
Sample a minibatch from dataset
7:
Train $ψ$ to minimize $L_{C E} (y_{t}, {\hat{y}}_{t}) - γ L_{C E} (y_{p}, {\hat{y}}_{p})$
8:
Train $θ_{a}$ to minimize $L_{C E} (y_{p}, {\hat{y}}_{p})$
9:
Train $θ_{t}$ to minimize $L_{C E} (y_{t}, {\hat{y}}_{t})$
10:
end for