Stealthy Poisoning Attacks Bypass Existing Defenses in Regression Settings

The researchers propose a form of optimal stealthy poisoning attack against regression models, considering varying degrees of detectability, and show that it can bypass existing state-of-the-art defenses. Regression models are widely used in industrial processes, engineering, and the natural and physical sciences, but their robustness to poisoning has received little attention previously, and existing studies often adopt unrealistic threat models. The paper also proposes an evaluation method based on target normalization to measure the trade-off between attack effectiveness and detectability, and develops a new defense, BayesClean, which outperforms previous defenses when attacks are stealthy and the number of poisoned points is large.