How to Detect Poisoned Data in Machine Learning Datasets