1: \begin{abstract}

2: We address the problem of minimizing a convex smooth function $f(x)$ over a compact polyhedral set $D$ given a stochastic zeroth-order constraint feedback model. This problem arises in safety-critical machine learning 

3: applications, such as personalized medicine and robotics. In such cases, one needs to ensure constraints  are satisfied while exploring the decision space to find optimum of the loss function. We propose a new variant of the Frank-Wolfe algorithm, which applies to the case of uncertain linear constraints. Using robust optimization, we provide the convergence rate of the algorithm while guaranteeing feasibility of all iterates, with high probability.\footnote{We thank the support of Swiss %... under the grant xxx (remove ack section)This work was gratefully supported by Swiss 

4: National Science Foundation, under the grant SNSF 200021\_172781, and ERC under the European Union's Horizon 2020 research and innovation programme 

5: grant agreement No 815943.}

6: \end{abstract}

7: