Fit the estimate (for example a median) on training data only, apply it to both splits, and keep a flag column so filled values are never mistaken for observations.
Related: Data leakage, Missing value (null)
Glossary
Imputation
Filling missing values with an estimate, flagged as filled.