Towards a Controllable and Reversible Privacy Protection System for Facial Images through Enhanced Multi-Factor Modifier Networks

. 2023 Feb 1;25(2):272. doi: 10.3390/e25020272

Algorithm 1. Multi-factor Modifier

Input : a set of face images I with face domain identity F a c e I D_{k}

and face attributes A_{k}

\forall k \in Φ

. Each target face (de - identified face) F_{p}

has one password P_{t}

, \forall t \in Φ

. Training iteration J_{k}

Output: Multi-factor modifier M and Discriminator V

fori $= 1 to J_{k}$ do
$Randomly select images I_{i}$ $with face domain identity F a c e I D_{i}$ and randomly $select image I_{t}$ $with face domain label F a c e I D_{t}$ $. Besides, F a c e I D_{i}$ is different $from F a c e I D_{t}$ $, \forall i, t \in Φ$
Update Discriminator V with the following objective function:
$L_{d i s c r i m i n a t o r} = L_{a d v} + L_{r e c_c l s}$
if (i + 1) mod 5 = 0 then
Update Modifier M with the following objective function:
$L_{f} = λ_{b} L_{b} + λ_{a t t r_c l s} L_{a t t r_c l s} + λ_{a d v} L_{a d v} + λ_{f e a t} L_{f e a t} + λ_{r e c} L_{rec} + λ_{d o m a i n_c l s} L_{d o m a i n_c l s}$

Where $L_{f}$ defined in Equation (12), respectively.
end if
end for
Repeat step 1 to step 10 until convergence
returnM and V