A Zeroth-Order Method for Adversarial Attacks

The diffusion of machine learning technologies and their applications is increasing nowadays, becoming more and more important in everyday life. One area of interest in this field is adversarial machine learning, where models are exposed to perturbed input data, known as "adversarial examples", to induce classification errors. This causes a serious threat to the reliability and security of models, especially when used in real-world applications. This thesis will investigate the generation of adversarial examples to attack deep neural networks used for image classification. The aim will be to minimise a black-box function using algorithms developed for structured optimisation. The formulation of the underlying mathematical model allows us to solve large-scale instances of the problem and find sparse solutions, generating adversarial examples that are very similar to the original images.