1
$\begingroup$

Is any classifier not subject to fooling as in here?

I agree that the question is related to the other one as commented by Philip. But I guess it is not completely a duplication as pointed out by hisairnessag3. What I wanted to ask is that any classifiers inherently do not subject (or less prone) to attack. I have a feeling that non-linear classifiers should be less susceptible to attack. Btw, any benchmark on say, simple K-nearest neighbor classifiers is available?

$\endgroup$
1
$\begingroup$

I would say this is not necessarily a duplicate but quite similar to some other questions. However, I will answer the question posed here.

At a theoretical level, what you are asking is there any algorithm that cannot be tricked into predicting the wrong class?

The answer is: No

It is analogous to asking whether there is a perfect architecture for an arbitrary classification problem, and that answer is quite obviously not. It would also not likely to be terribly difficult to show this is provably the case, at least for a class of algorithms(i.e connectionist models).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.