Worst-Case Morphs Using Wasserstein ALI and Improved MIPGAN

Algorithm 1

Our training procedure.

repeat

           Draw samples from the dataset

               Draw random latent emb.

         Get real embeddings

        Generate fake images

       Critic output for real data

      Critic output for fake data

    Reconstruct real images

  Translate batch to                          get pairs for morphing

       Generate morphs

           Get FR embeddings

       Get worst-case emb.

       Compute pixel loss

          Compute FFL loss

                           Compute FR-based losses

         Compute critic loss

                            Compute generator loss

             Gradient update on critic

        Gradient update on encoder

        Gradient update on decoder

until convergence