Craft Adversarial Data
Crafting adversarial data is like creating a 'magic' image or sound that looks or sounds perfectly normal to a human, but tricks an AI into making a mistake, like misidentifying a stop sign as a speed limit sign.
Crafting adversarial data involves applying specific, often subtle, perturbations to input data to force an AI model to produce an incorrect output, such as a misclassification or a missed detection, while often maintaining the original semantic meaning of the input to human observers.
Crafting adversarial data is the systematic generation of inputs, typically via optimization-based or heuristic methods, designed to exploit vulnerabilities in an AI model's decision boundaries. These inputs are modified—often under constraints such as L-norm bounds to ensure human imperceptibility—to induce a specific adversarial objective, ranging from targeted misclassification and evasion to the degradation of model integrity or resource exhaustion, contingent upon the adversary's level of model access (white-box vs. black-box).