How does Naïve Bayes perform under adversarial feature poisoning attacks?

Naïve Bayes is vulnerable to feature poisoning because likelihood estimates can be shifted by maliciously injected samples.

Adversarial poisoning introduces carefully crafted samples to manipulate class-conditional distributions. Since NB relies on frequency-based estimation, even small amounts of poisoned data can skew P(x|C) significantly for rare features. Smoothing reduces but does not eliminate this vulnerability.