On security and sparsity of linear classifiers for adversarial settings