Phenomenon where test or validation dataset information improperly influences machine learning model training, causing overfitting and artificially high performance metrics. Includes split contamination, feature leakage, temporal leakage, and prevention techniques to ensure correct model generalization.