This could work, but it is important to note that a lot of ML algorithms trained in a closed domain (no "other" class) will be pretty bad at knowing what they don't know. This is an open problem in ML.