Robust Learning of Optimal Auctions