Batch normalization amplifies memorization of outliers and increases susceptibility to membership inference attacks, posing a privacy risk to models using this technique. Researchers conducted an empirical study on multiple datasets and architectures, observing a higher memorization of outliers and increased susceptibility to attacks.