A penetration tester discovers that by appending a specific sequence of characters to any input, a content moderation AI model consistently fails to classify harmful content. This is an example of which AI security vulnerability?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Here's the deal — b is correct because a universal adversarial perturbation is a fixed pattern that, when added to any input, consistently causes a model to misclassify it. The key characteristic is that the same perturbation works across diverse inputs universally.
Full explanation below image
Full Explanation
B is correct because a universal adversarial perturbation is a fixed pattern that, when added to any input, consistently causes a model to misclassify it. The key characteristic is that the same perturbation works across diverse inputs universally. A backdoor (A) requires a specific trigger embedded during training and would typically be domain-specific, not a universally applicable character sequence for all inputs. Model extraction (C) involves learning decision boundaries through queries. Membership inference (D) tests whether data was used in training.