Why stop_queries uses working_ind to update itself before updating working_ind? about rays HOT 3 CLOSED

stop_queries aims to measure the number of queries needed until the first time we found a successful attack (note that we can still continue performing RayS to get better decision boundary estimates), so it is different from the final number of queries. In this regard, we update the stop_queries for all working_ind each time and stop updating them once they are not in working_ind (success attack found). Hope this answers your question.

from rays.

@uclaml Why you place working_ind = (dist > self.epsilon).nonzero().flatten() after stop_queries[working_ind] = self.queries[working_ind]??? Isn't it to be placed before stop_queries[working_ind] = self.queries[working_ind]?

from rays.

A simple counterexample: assume it works in the order you said. stop_queries is initialized as a zero vector and working_ind is an all-one vector. Suppose in the first attempt we successfully attacked all indices, meaning dist<=epsilon for all indices. Now you will first update working_ind to be an empty list and therefore there will be no update on stop_queries (remains zero vector). Yet this is not correct, as the stop_queries should be the queries used in the first attempt. Hope this helps.

from rays.