Adversarial testing · Question 37 of 60
Which assumption underlies transferring adversarial inputs from an equivalent model to the original model?
1 answer is correct. Medium
Show answer
Answer
- D Equivalent models share classification boundaries (transferability)
Explanation
Behind this is transferability: equivalent models draw their class boundaries in similar places. What fools the known substitute model therefore often fools the original too, so these inputs are tried out there.