1: \begin{abstract}

2: In this paper, we study the problem of constrained  min-max optimization {in a black-box setting}, where  

3: the desired optimizer cannot access the gradients of the objective function but may query its values.  We present a principled optimization framework,  integrating a  

4: zeroth-order (ZO) gradient estimator with 

5: an alternating projected stochastic gradient descent-ascent method, where the former only requires a small number of function queries and the later   needs  just one-step descent/ascent update.  

6: We show that the proposed framework,  referred to as \textit{ZO-Min-Max}, has a sub-linear convergence rate under mild conditions and scales  gracefully with problem size.

7: We also

8: explore a promising connection between black-box min-max optimization and  black-box evasion and poisoning attacks in adversarial machine learning (ML). Our empirical evaluations on these use cases demonstrate the effectiveness of our approach and its scalability to dimensions that prohibit using 

9: recent  black-box solvers. 

10: \end{abstract}

11: