[12] OpenSMax: Unknown Domain Generation Algorithm Detection. ECAI 2020: 1850-1857

This article is about openmax Improved version , It is also through the post-processing of the cross entropy classification network , Used for open set domain name detection （ Unknown domain name discovery ）. openmax The reason is that the established single class model is not an extremum machine , It is ocsvm, And modeling is not just penultimate layer , It also includes the last layer

In training , Just use ordinary cross entropy .

At testing time , For each correct classification sample activation value of each class, establish a ocsvm（ Algorithm 1,）, And used to correct a new K+1 Class classification probability （ Algorithm 2）. meanwhile , Also set a threshold , If the maximum class correction probability of the sample is less than the threshold , Classified as unknown , Otherwise, it is classified as the maximum probability class .（ Here, the maximum probability class itself may also be an unknown class ）