Stealthy Poisoning Attacks Bypass Defenses in Regression Settings

Researchers have developed new stealthy poisoning attacks that can bypass current defenses in regression models used across industrial and scientific applications. The study introduces BayesClean, a novel defense mechanism that better protects against these sophisticated attacks when poisoning attempts are significant.

• →Current regression model defenses are vulnerable to sophisticated stealthy poisoning attacks that can evade detection.
• →Researchers developed a new attack formulation that considers different degrees of detectability to bypass state-of-the-art defenses.
• →A new methodology using normalization of objectives allows better evaluation of trade-offs between attack effectiveness and detectability.
• →BayesClean defense mechanism shows improved protection against stealthy attacks compared to previous methods.
• →The research highlights significant security gaps in regression models used in industrial processes and scientific applications.