On Optimal Learning Under Targeted Data Poisoning