Closed-source machine learning models can be evaluated for private data leakage by issuing targeted queries through application programming interfaces