Sensitivity of Bayesian Inference to Data Deletion and Replication
Adversarial machine learning (AML) has shown that statistical models are vulnerable to data manipulation, yet most studies focus on classical methods. We extend white-box poisoning attacks to Bayesian inference, demonstrating its susceptibility to strategic data manipulations. Our attacks, based on selective deletion and replication of observations, can steer the Bayesian posterior toward a desired distribution—even without an analytical posterior form.
We establish their theoretical properties and empirically validate them in synthetic and real-world scenarios. Interestingly, in some cases, modifying a small fraction of carefully chosen data points leads to drastic shifts in inference.
Keywords: Bayesian inference MCMC adversarial machine learning