White-Box Optimization
White-box optimization is a method where a hacker can see exactly how an AI 'thinks' and use that knowledge to create custom inputs that trick the AI into making mistakes.
In White-Box Optimization, the adversary leverages full visibility into the target model's architecture, weights, and gradients to perform iterative optimization, resulting in highly effective adversarial perturbations tailored to that specific model.
White-Box Optimization refers to an adversarial attack paradigm where the adversary possesses complete knowledge of the target model's parameters and architecture, enabling the direct computation of gradients with respect to the input to minimize a loss function, thereby generating adversarial examples that are maximally effective against the target model.