What is feature leakage in Keras model training?

Feature leakage occurs when training data contains information that would not be available at inference time.

This leads to artificially high training accuracy but poor generalization. It often happens due to incorrect preprocessing, target encoding, or time-based data leakage.