Which statement defines Model Evasion?

• A. Attacker modifies data to circumvent the model's correct prediction, causing misclassification.
• B. The model outputs are inverted to reveal sensitive information.
• C. The attacker steals training data by querying the model.
• D. The model generates highly realistic but false content.

Answer: (A)

Model evasion happens when an attacker crafts inputs so the model makes incorrect predictions at inference time, without changing the model itself. This matches the statement that describes an attacker modifying data to bypass the model’s correct prediction and cause misclassification. The attack leverages weaknesses in how the model separates classes around its decision boundary, often using carefully perturbed inputs that look normal to humans but fool the model.

This is distinct from other security concerns: revealing sensitive information from outputs (model inversion), stealing training data by querying the model (model extraction or data theft), or generating highly realistic but false content (synthetic content generation). Those are different attack types with different goals and mechanisms.