Abstract: Existing research has demonstrated that deep neural networks are susceptible to model extraction attacks, where an attacker can construct a substitute model with similar functionality to the ...