Algorithm 1: Image de-identification content generation.

Input: The private image $I_p\in {\Re }^{n\times m\times 3}$ to de-identify; A pretrained generator $G(·)$.   Output: The de-identified image $I_d$ optimized via $G(·)$   Initialize latent vector $\omega$, differential privacy Laplace noise with $\mathsf{\Delta }f$ and $ϵ$;   while not converged do $| I_p\simeq I_p^{\prime }=G\left({\omega }^{*}\right)$;   end$\phantom{(}$ $I_{pd}=G({\omega }^{*}+Lap\left(\frac{\mathsf{\Delta }f}{ϵ}\right))$ ;