Conference paperTowards verifying robustness of neural networks against a family of semantic perturbations