Positive label smoothing increases model privacy leakage, while negative label smoothing offers a defense against model inversion attacks.
Label smoothing can enhance model privacy leakage in model inversion attacks, but negative label smoothing offers a defense strategy with a better utility-privacy trade-off.