This paper proposes a novel algorithm, JMA, which performs targeted attacks through Jacobian-based Mahalanobis distance minimization to overcome the limitations of existing targeted attack methods that focus on a single class. JMA considers the effort required to move the latent space representation in a specific direction from the input space and transforms it into a non-negative least squares (NNLS) problem using Wolfe's duality theorem to find an optimal solution. This provides an optimal solution to the linearized version of the adversarial example problem proposed by Szegedy et al. It is effective in various output encoding schemes and, in particular, has the advantage of being able to target and modify up to half of the labels in multi-label classification. It also operates with a smaller number of iterations than existing methods.
